APMIS 96: 379-394, 1988

Some new, simple and efficient stereological methods and their use in pathological research and diagnosis
Review article
H. J. G. GUNDERSEN, T. F. BENDTSEN, L. KORBO1, N. MARCUSSEN, A. MLLER1, K. NIELSEN2, J. R. NYENGAARD, B. PAKKENBERG1, F. B. SRENSEN, A. VESTERBY3 and M. J. WEST Stereological Research Laboratory, University Institute of Pathology and 2nd University Clinic of Internal Medicine, Institute of Experimental Clinical Research, University of rhus, Neurological Research Laboratory1, Hvidovre Hospital, Institute of Pathology2, University of Copenhagen, Rigshospitalet, University Institute of Pathology3, rhus Amtsygehus, Denmark

_ _

Gundersen, H. J. G., Bendtsen, T. F., Korbo, L., Marcussen, N., Mller, A., Nielsen, K., Nyengaard, J. R., Pakkenberg, B., Srensen, F. B., Vesterby, A. & West, M. J. Some new, simple and efficient stereological methods and their use in pathological research and diagnosis. APMIS 96: 379-394, 1988. Stereology is a set of simple and efficient methods for quantitation of three-dimensional microscopic structures which is specifically tuned to provide reliable data from sections. Within the last few years, a number of new methods has been developed which are of special interest to pathologists. Methods for estimating the volume, surface area and length of any structure are described in this review. The principles on which stereology is based and the necessary sampling procedures are described and illustrated with examples. The necessary equipment, the measurements, and the calculations are invariably simple and easy. Key words: Cavalieri's principle; isotropic sections; star volume; surface area; stereology; vertical sections; volume. H. J. G. Gundersen, Stereological Research Laboratory, University Bartholin Building, DK-8000 rhus C, Denmark.

THE BASIC IDEAS AND TOOLS OF STEREOLOGY

At a practical level, stereological methods are precise tools for obtaining quantitative information about three-dimensional, microscopic structures, based mainly on observations made on sections. Two-dimensional sections contain quantitative information about three-dimensional structures only in a statistical sense. For this statistical information to be "true" or unbiased a few requirements must be fulfilled about the sections and the way they are made. In practice it is nearly always very easy to fulfil these requirements (often as easy as doing it the wrong way!) and throughout the text the sampling methods are described and references are given to full descriptions of the techniques.

Two concepts are often referred to in this and other papers dealing with stereology: "Unbiased" and "efficient". The concepts are used in a statistical sense and mean "without systematic deviation from the true value" and "with a low variability after spending a moderate amount of time", respectively. In the following, two-dimensional information and the way to obtain it effeciently are dealt with first, then the basic methods for measuring volume, surface area, and length are described. For discrete particles like cells, nuclei, glomeruli etc., a number of new methods are available for measuring their number and mean sizes in a very direct way. These methods are described in the second, forthcoming part of the review. Al1 the methods are used extensively in biological research and pathology, as indicated by a number of references
379

APMIS 96: 379-394, 1988

to recent studies throughtout the text. Hopefully, the examples will make many pathologists appreciate that, especially in their own interesting tissue, effiicient and unbiased quantitative data may mean the difference between "interesting observations" and real knowledge.
TWO-DIMENSIONAL QUANTITATION

The two-dimensional methods also serve as an introduction to the full set of stereological methods, since the same principles are involved.

The strength of stereological methods is that they report data for three-dimensional structures in terms of three-dimensional quantities - not just in terms of the two-dimensional sections which are most often used for the observations. The size of a cell may naturally be expressed in m3, not by the area in m2 which its profile happens to occupy on a section. Nevertheless, there are many situations where simple, fast and reliable methods for quantitation of two-dimensional structures are needed.

The Number of Profiles per Area In Fig. 1 is shown a section seen through a "window": a photograph, a field of vision or anything else which imposes artificial boundaries on the visible part of the section: How many profiles do we see? This is not as simple a question as it may seem: There are 31 profiles completely inside the window, but, in addition, 10 profile parts are seen on the edges (they may or may not represent 10 different profiles, but we have no way of knowing that). So, what is the number of profiles; 31, 41, 36 or what? Most biologists have learned how to count erythrocytes inside countingchambers; i.e. counting on certain edges and corners of the counting frame, but since it systematically overestimates the number, it is a biased rule, see the review by Gundersen (1978). All known practical counting rules more than ~ 10 years old are biased. For humans, (machines are different), the only known unbiased frame and its counting rule is also illustrated in Fig. 1: In addition to profiles completely inside the frame one counts all profiles with anything inside the frame provided they do not in any way touch or intersect the full drawn exlusion edges or their extensions (Gun-dersen 1977). The answer to the question is 27 profiles. This number is not directly comparable to any of the above figures because the area of the frame is less than that of the window. For irregular profiles especially there must be a "guard-area" around the frame for counting correctly. The natural way to report such a number is as a numerical density in the plane, i.e. the number of profiles per area, a quantity which in stereology is given the symbol Q A
2 # of profiles prof Q (prof/sect) = Q((sect )) = areaof frame = 272 = 3.1u A 8.8u A

Fig. 1. Above. A restricted field of vision of a section with some profiles. Hatched profiles are completely inside the field. Below. The same field with an unbiased counting frame (Gundersen 1977), the hatched profiles are counted according to the counting rule described in the above reference and in the text.

assuming that the area of the counting frame is 8.8 in some recognizable areal units, u2. The main advantage of reporting the number of profiles in such a standardized way - that it makes it possible to draw comparisons between different techniques and observers - is so great that pathologists always quantify e.g. "cellularity" in terms of number of cell profiles per unit area and never as just cells or mitoses per field of unknown size and magnifica-

380

RECENT STEREOLQGICAL DEVELOPMENTS FOR PATHOLOGY

Fig. 2. The same field as in Fig. 1, now superposed with

a test system with a set of regularly spaced points (and a counting frame). For estimates of the areal fraction of profiles or mean profile area one counts all points hitting profiles (disregarding their relationship to the frame). A point is considered a hit if a profile (including its boundary) covers the upper right corner in the cross where the two lines cross each other.

not to mention the capital investment in hardware. It takes ~ 18 seconds to count the hits in Fig. 2. To trace the boundaries of all profiles manually without a large bias takes 5 to ten times more time, without any gain in real precision (no automatic image analyzers can work on ordinary biological tissue), for examples of direct comparisons see for example Gundersen et al. 1981, Regeur & Pakkenberg 1988, and references therein. For serious use one must again sample the fields systematically or by some other random mechanism and for just 5 such fields quantified in this way the coefficient of SEM error, CM = MEAN of the estimate is in the order of 5%. A nomogram for reading off the precision of point-counting estimation of areas is found in Fig. 18 in Gundersen & Jensen 1987. As it is well known, Pp(struct/sect) may also be used as an estimate of the volume fraction Vv(struct/specimen) of the structure under study, described in the next section. One must then as well fulfil some further requirements with respect to the sampling of the sections inside the specimen. Note that the arrangement of the points is irrelevant to the estimator and that the magnification need not be known because we are only estimating a relative area. Later on we shall see another use of a test system for estimating absolute areas where we must specify the above information about the points somewhat more precisely. Another question which a pathologist would often ask is the area of a typical profile, i.e. the mean profile area a(prof). Since the relative area and the relative number of profiles may be estimated by the above techniques, this is a very simple problem to solve, see also Fig. 2:
a(prof) =
A 8 prof / sec t ) A q ( prof / sec t ) a

tion. For the number of profiles per area to be of any use one must, however, also specify how the fields were sampled. If the fields were sampled just because the pathologist liked them the best the number of profiles per area may well characterize the pathologist more than the section. The simplest sampling system which also fulfils the statistical requirements for a scientific study is one where the fields are sampled systematically and independent of their content - and of the observer - but all inside a well defined region.

The Area of Profiles The simplest way of reporting "how much there is" of the structure in the profiles shown in Figs. 1 to 3 is to quantify the areal fraction of the structure, generally denoted AA
A (struct/sect) = P(sect ) = totalareaof structure= 18 = 0.23 total of sections area 80 A
P( struct )

relative area of profiles relative# of profiles

18 80 27 8.8 u 2

= 0.073u

A quantity which can be estimated by counting points which hit the structure, P(struct), divided by points hitting the section, here illustrated by the example in Fig. 2. Note that in the simple situation where the section fills the whole window one need not count the points hitting the section, all 80 points in the test system do that. In terms of time and effort there is no way to do it more efficiently,

Note that the straightforward estimator is the ratio of two independent estimators: the points hitting profiles are counted irrespective of whether the profile in question is counted in the frame or not, see Fig. 2. It is only if we need to know the distribution of sectional profile areas that the counting of points is restricted to profiles sampled in the frame (and the point density should then be somewhat higher).

The Boundary of profiles Another way of quantifying the structures seen in a microscopic field is to measure the length of their
381

GUNDERSEN et al

profile boundary b (prof) just as before. When estimating the boundary length on sections one must ensure that either the boundary is globally isotropic, i.e. without any detectable preferred orientation, or that the test system or the fields of vision are rotated at random. As it is well known, 2 IL(prof/sect) is an estimator of surface density, SV(structure/specimen), but only when the orientation destribution of the sections is also taken into account, a problem dealt with later on.

Fig. 3. The same field as in previous figures, the test

system now includes a set of test lines: the upper edge of the lines joining some crosses. The length of one test line is the distance between two points with the definition given in Fig. 2.

boundaries. Again, it is far too slow an operation to trace the outline of all profiles for that purpose, the setup shown in Fig. 3 is much more efficient. A test system with lines of known length is part of the integral test system. This only means that the lines and the points - and the frame, if there is one - are inseparable because they are drawn together in the same physical test system, which is a piece of paper or a transparency, Jensen & Gundersen 1982. The arrangement of the different parts of the system is irrelevant, much to the surprise of most investigators. In the integral test system the ratio between the total length of test lines and the number of test points is l/p u, i.e. the test line length in length units u per test point, simply. Whenever a test line intersects a profile boundary it thereby makes an intersection point l(prof). In addition one must know the number of test points P(sect) hitting the section like before. The relative profile boundary length or profile boundary length per area of section BA is in terms of the example in Fig. 3
B = A =
2

Test Systems and Efficiency Both two- and three-dimensional quantitation has suffered a lot from the very widespread misunderstanding that one has to count thousands of points - but this is never the case. With respect to the precision of the estimate the most important aspect is to sample enough fields of vision to climate any variation from field to field. If that is done, the necessary total number of events counted need never exceed 100 to 200, i.e. for the estimation of a(prof) one counts roughly 100 P(prof) and 100 Q(prof) over the set of fields sampled in one specimen, cf. Gundersen et al. 1980, Gundersen & sterby 1981, Gundersen 1986, and the numerous references in the latter review. It is also possible to predict the optimal

I
18 80

(prof/sect = 0.86u

)=
-1

p 21

profile int er sec tions po int s on sec tion

2 0 . 41

Note that we may estimate the total length of test line on the section by counting points on the section since l/p = 0.41 u, the test line length per point in the test system, is known. With the above estimate of the relative boundary at hand one may then estimate the mean

Fig. 4. A test system with three sets of points and two sets of test lines, arranged in a regular tessellation (a complete covering of the plane) of the unit enclosd in stippled lines with area u2. The three point sets are: all encircled points (one per unit), all points at the ends of test lines (four per unit), and all points not on test lines ( 16 per unit). The two line sets are all lines with an encircled point at the end (one per unit) and all lines (two per unit).

382

RECENT STEREOLQGICAL DEVELOPMENTS FOR PATHOLOGY

sampling scheme from data of the study itself. Such a prediction may even be read off in a simple nomogram, see Gundersen et al. 1980 and Gundersen & sterby 1981. Test systems are usually so simple that the easiest way to get one is to draw it. For many purposes two or three basic designs are suffiicient. For a new study one then just makes a copy of a suitable test system at a magnification where the above sampling intensity is realized, see for example Gundersen 1984. In many cases it is a great advantage to have more than one set of points and lines in the same test system. One then uses the few, "coarse" points for counting on the section and another set of many, "fine" points for counting on something which is only covering a fraction of the section. The known ratio between the point sets is then taken into account in the calculations. One of the most popular test systems in our laboratories over a number of years is shown in Fig. 4.

Fig. 5 . A coronal section of a human brain with a transparent test system superposed. For the Cavalieri-estimate of total cortex volume one counts all encircled test points which hit the brain cortex sectioned by the upper section plane, disregarding oblique pial surface visible below the section plane. See text for further details. The length of one test line is 2 cm.

CAVALIERI'S DIRECT ESTIMATOR OF THE VOLUME OF ANYTHING FROM SECTIONS

This is the first of some examples of very simple stereological estimation procedures which have turned out to be very useful in a vast number of situations. They are also characterized by their strong property of unbiasedness, i.e. in order to use them one does not have to make any unrealistic assumptions about the structure with respect to shape, orientation etc., etc. The irregular shape of various organs in the human body has imposed an interesting and in many cases unsolved problem for estimation of reliable volumes, the human brain cortex being an example in point with the folded outline of its surface. With the introduction of stereological methods some 20 years ago and a growing improvement of techniques since then, volume estimation is no longer a problem. In the following, a method for estimation of brain cortex, white matter, central grey structures and brain ventricles is presented. The principle is evidently extremely general and therefore applicable to almost anything, the mathematical basis of the estimator is very old. For all intents and purposes it was given by the Italian mathematician Cavalieri who lived from 1598 to 1647. Cavalieri showed that the volume of any object V(obj) may be estimated from parallel sections separated by a known distance t, by summing up the areas of all cross sections of the object a(prof) and multiplying this figure by t: V(obj) = t a(proit) There are no other conditions except that the position of the first section must be random in the object, see Gundersen & Jensen 1987. Specifically, the estimate is completely independent of the orientation of the set of sections and of the shape of the object.

Method The fixed brains are embedded in 20% gelatin in 2% aqueous phenol and stored cold for seven days after which surplus gelatin is trimmed off (alternalively, the brains may be embedded in 7% Agar). The brains are then cut into parallel, coronal slices six mm thick. Since the brain may not be sliced with a precise distance of six mm the average slice thickness t can be estimated by measuring the length of the brain omitting the first and last slice and then divide with the total number of slices minus 2. On all cut surfaces on all sections one then applies at random a test system with regularly arranged points, see Fig. 5. For each structure under study one now simply counts all points P(struct) which hit it. The absolute area of the cross section of the structure is
383

GUNDERSEN et al.

a(sect) = a(p) P(struct) u2 where a(p) is the area in known units u2 associated with each point in the regular test system, see Fig. 4. The total volume of each structure: cortex, white matter, central grey matter, ventricles etc., is now estimated after adding all points hitting that structure V(struct) = t a(p) P(struct) u3 The remarkable freedom of the estimator from assumptions regarding shape and orientation has already been commented upon. It is also a very useful estimator of the volume of very irregularly shaped cells (for a range of examples, see Gundersen & Jensen 1987), where one never has to resort to serial reconstruction, a very tedious procedure from which one gains no reliable, quantitative information. Another noticeable feature of the estimator is its effiiciency, studied in detail in Gundersen & Jensen 1987. Even for a very irregular three-dimensional structure like the human brain cortex a total of ~ 200 points counted on 10 to 15 sections will provide an unbiased estimate of the true volume with a precision better than 5%, a procedure which takes ~ 15 min. For a number of further practical details see Pakkenberg 1987 and Regeur & Pakkenberg 1988 wherein large series of human brains are studied. Most of the stereological estimators of structural quantities described in the literature, see for example Weibel's book from 1980 or the reviews by Gundersen 1980 and 1986, are two step procecures: One estimates in the first step the density of the structure of interest in the volume of the containing or "reference" space V(ref). All these densities: Volume per volume VV, surface area per volume SV, length per volume LV, or number per volume NV, are ratios which generally do not allow one to make conclusions about changes in the absolute amount of the structure under study (some spectacular examples of mistakes made over the years are discussed in the section entitled "The Reference Trap" in Brndgaard & Gundersen 1986). The greatest value of Cavalieri's estimator is therefore that it effortlessly allows one in the second step of an ordinary stereological estimation procedure to obtain estimates of the volume of the reference space even if it is enclosed in some other tissue. For an isolated organ like the
384

thyroid gland or the liver with a specific gravity ~ 1.0 one can in most cases simply weigh it to get an essentially unbiased estimate of its volume. The absolute amount of the structure is then the density times the absolute volume of the reference space. As already mentioned, the general estimator of volume fraction is PP(struct/ref) = VV(struct/ref) for the use of which the sections must have random positions in the specimen, in practice a systematically random set of roughly equidistant sections is likely to be the most efficient sampling scheme. Most of the practical details of this estimator have already been dealt with. It is, however, worth mentioning again that for almost all specimens five systematic sections and a total of < 200 points counted will provide a very precise estimate, see several of the references already provided.

ESTIMATION OF LENGTH OF FIBRES, TUBULES, CAPILLARIES ETC.

The probability that a given structure is hit by a randomly positioned and randomly oriented section is proportional to its linear dimension or length. It follows that if we observe how often a particular type of structure is hit on a section - i.e. count the number of profiles - we have a simple and direct estimator of its total length (if the volume of its containing space is known). It turns out to be as simple as L(struct) = 2 QA(struct/ref) V(ref) u where the numerical density in the plane QA is estimated as described above for number of profiles in general. From a few EM-sections through randomly sampled glomeruli (sterby & Gundersen 1978) and en estimate of the total volume of glomeruli in a rat kidney of ~ 30 mm3 we may then learn that the length of capillaries in rat glomeruli in one kidney is (sterby & Gundersen 1980) L(cap) = 2 QA(struct/ref) V(ref) = 2 11,000 30mm ~ 600m

RECENT STEREOLQGICAL DEVELOPMENTS FOR PATHOLOGY

and that this length is increased by 200 m after just 4 days of diabetes! There is a rather strong requirement to fulfil in the above estimate, namely that the glomerular capillary profiles were counted on isotropic EMsections. This is a practical problem to which a very simple and strong solution without any assumptions has recently been found and is described below. There is also the mild assumption that the structure is "much longer than it is wide", i.e. its length is much larger than its diameter, see Mattfeldt et al. 1985 and sterby et al. 1988. Contrary to general belief it is not in any way a requirement that the structure should be straight or cylindrical, one still counts just the number of isolated profiles. It follows that for a tubular structure one may choose to count the isolated inner profiles of the lumen or the isolated outer profiles. On a given section their number (counted in an unbiased frame) will generally not be the same, but averaged over a suitable number of fields they estimate the same length, see Seyer-Hansen et al. 1980 for an illustrative example in which kidney tubule length was estimated.

THE SIMPLE WAY OF MAKING A TRULY ISOTROPIC SECTION: THE ORIENTATOR

Isotropic sections are necessary for a number of stereological procedures: The above mentioned estimator of length, the very efficient estimator of membrane thickness and its distribution ( Jensen et al. 1979), and some of the newer estimators to be described later. The procedure is named the Orientator and is described by Mattfeldt 1988. Fig. 6 illustrates the practical steps with an example. More generally, one might embed the specimen in Agar and then make the two consecutive sections illustrated in Fig. 6. If the structure under study is highly anisotropic the estimation may be carried out on two more section planes perpendicular to the last section plane in Fig. 6 and also perpendicular to each other, a co-called Orthrip (Mattfeldt et al. 1985). It takes one minute to make the two sections for the Orientator, hardly an excuse for ignoring structural anisotropy. For a procedure for estimating the number of capillaries in glomeruli, also relying on isotropy, see Nyenguardet al. 1988.

Fig. 6. The Orientator by Mattieldt 1988. Above. The

specimen is put roughly at the centre of the circle with equidistant divisions along the perimeter. The specimen must be placed on a previously cut, arbitrary surface (which might be through some embedding material). A random number among the divisions is selected (using a random number table). The specimen is cut with a section perpendicular to the plane of the circle and in the selected direction (the 14 to 32 direction in the example). The cut surface and the bottom surface makes a straight edge. Below. The specimen is placed on the surface just cut with the straight edge parallel to the 0 to 0 direction of the second circle with non-equidistant divisions along the perimeter. The divisions are produced as shown in Fig. 16. A new random number is selected and the specimen is cut in this direction (65 in the figure) with a section perpendicular to the circle. The last surface cut is isotropic uniform, IU, in the specimen, i.e. without regard to the position of the specimen in the first step the last surface has an orientation which is taken from all possible orientations with constant (uniform) probability.

383

GUNDERSEN et al

Fig. 7. a. Simple sampling scheme for vertical sections. An arbitrarily shaped object is sliced at random into a number of slabs with the same thickness, all parallel to a horizontal plane chosen a priori. A systematic set of three slabs are selected, all perpendicular to the vertical axia (arrow). b. A horizontal slab, seen from above, is cut into vertical sections in three systematic directions. The first direction around the vertical axis is selected at random and the other directions can be determined using a wax plate with a stellate pattern, as shown. A set of vertical sections have been obtained.

ESTIMATION OF SURFACE AREA OF ANYTHING FROM SECTIONS On a section, the surface of a structure shows up as a boundary. There exists a relationship between the length of the boundary seen and the surface area of the structure, but this is not an efficient estimator because all the boundaries must be traced. Instead, one counts intersection points I between the boundary and so-called IUR test lines: S(struct) = 2 IL(struct/ref) V(ref)

VERTICAL SECTIONS, THE EFFICIENT SOLUTION TO MOST ANISOTROPY PROBLEMS Most existing stereological procedures require some kind of isotropy. Plane sections of the structure in question must be isotropic, uniform random (IUR) planes, or the structure itself must be statistically isotropic. The only two exceptions from this are estimation of volume and number. These basic requirements give the pathologist, working in the practical field of stereology, serious problems, due to the fact, that biological structures are often anisotropic. In such situations it may be diffiicult or impossible to make IUR tissue sections. Indeed, the pathologist often prefers to take sections at a particular orientation to obtain information which is absent on other sections. A practical solution to the problem of anisotropy is presented in this section, i.e. Vertical Sections. General remarks on the basic principle of vertical sections, its impact on sampling procedures and how to make measurements on vertical sections wil1 be reviewed. An example which shows how to estimate surface area from vertical tissue sections is briefly presented. Definition of Vertical Sections A vertical tissue section is a plane section perpendicular to a given "horizontal" plane. The meaning of a horizontal plane is only a plane of reference,

2p I

I(struct) V(ref)u 2 P(ref)

in some analogy to the estimator of profile boundary. Considering the multitude of biological processes related to surfaces the estimator ranks among the most important of the stereological estimators. IUR in the above means isotropic, uniform random, i.e. the test lines must have isotropic orientation (all directions equally likely) and random position in three-dimensional space, not just on the section. It is possible to obtain this on isotropic sections, but on such sections one tends to loose track of the architecture of the tissue. However, a recently discovered method can produce isotropic test lines while still maintaining a preferred direction of sectioning.

386

GUNDERSEN et al

Fig. 8. Sampling scheme, with improved efficiency, for vertical sections. Consider the three slabs in Fig. 7a, as a set of parallel, systematic sections. Each slab is placed on the table. a. The firct slab is sliced into parallel "chips" of the same thickness in an arbitrary direction. Likewise, the next slabs are cut into chips, however, in another systematically selected direction obtained by rotating the slabs lying flat on the table. b. Parallel, equidistantly spaced chips are systematically selected (three in this example) for further processing. Now the vertical axis is defined for these three chips as running longitudinally through the chips. Then the chips are allowed to rotate freely around their vertical axis. c. Embedding of the chips has now been accomplished at random with respect to the rotation around the vertical axis (e.g. in Agar or paraffin). Sections from the block then provide a material, where the vertical axis is easy to identify (the long axis of the chips). Taking another set of chips at the stage shown in b, another vertical direction in the specimen will be defined, and thus, eventually a set of sections with a high number of vertical axes is obtained! for highly efficient stereological estimation.

which defines the orientation of the section. This is a rather handy definition, since the horizontal plane can be given by the tissue itself or generated artificially by the observer. Examples of tissues naturally possessing a horizontal plane are for instance tissues covered with flat epithelium like skin. The possibility to produce a horizontal, reference plane of any organ of arbitrary shape makes the principle of vertical sections applicable to every experimental setting.

Sampling Procedures for Vertical Sections In the practical handling of tissue for use in vertical section stereology, four requirements must be fulfilled, see Baddeley et al. 1986. The first three deals with the cutting of tissue specimens and are discussed here, whereas the fourth requirement concerns the stereological test system employed for the measurements. The number of steps involved in the handling of specimens may vary considerably in different applications. Requirement 1: Either the tissue must intrinsically possess an identifiable directional (vertical) axis, or the pathologist must generate such a direction. The axis only needs to be identifiable, and no symmetry with respect to the biological structure is required. It is very easy to define a vertical axis in muscle at the macroscopical level. In skin the vertical axis can be defined by the normal to the horizontal, reference plane represented by the macroscopical cutaneous surface. Tubular organs like gut and large blood vessels may be opened along their axis, flattened and thus made into flat specimens with an identifiable horizontal plane

(see Fig. 4 in Baddeley et al. 1986). The effiiciency of the method is greatest, when the vertical sections contain the preferred directional axis of the biological structure in question. However, the pathologist is completely free to generate an artificial vertical axis, which may be of special interest in the experiment. If an artificial vertical axis is desired, the arbitrary tissue specimen is sliced in several planes, all parallel to an arbitrary starting plane of arbitrary direction, i.e. the horizontal base plane. Vertical sections can then be sliced normal to the horizontal plane, see Fig. 7. The "horizontal slabs" should be of the same thickness, to make the

Fig. 9. "Object rotator,', which can be mounted on the object table of an ordinary miscoscope. After the histological specimen has been attached, the instrument enables free rotation of the specimen in the section plane and alignment with the vertical axis is easy. (Proto type manufactured bv Olvmous. Denmark).

387

GUNDERSEN et al.

where randomness is required. A set of systematic randomly positioned sections can be sliced normal to the horizontal plane by a semiautomatic tissue slicer, see Fig. 6 in Baddeley et al. 1986. Systematic random orientation is achieved by selecting the first direction at random, and the succeeding directions systematically with respect to the first, see Fig. 7. The pathologist has at this stage no freedom to select an especially advantageous direction. This freedom was exhausted at the first stage when the vertical direction was defined.
Fig. 10. A segment of a cycloid arc with orientation distribution proportional to the sine of the angle to the vertical axis. The length along the test curve is twice its height (parametric equation of a cycloid: x = -sin, y = 1-cos, where 0 < .

estimation procedure most efficient. Another design for generating vertical sections is shown in Fig. 8. where optimal conditions to increase efficiency are given. Requirement 2: All the vertical sections must be parallel to the vertical, i.e. normal to the horizontal, and the vertical direction must be identified in each section. In tissue sections showing a natural orientation, the vertical axis is easily identified. If an artificial vertical axis has been generated, as described above, this requirement can be satisfied by taking all vertical sections normal to the common horizontal plane, and including the lower face of the slabs in the sections. The vertical direction is then always perpendicular to the lower edge of the slab as seen on the vertical histological section. However, the vertical sections need to be truly normal to the horizontal plane, and care has to be taken to identify the horizontal plane during the processing of the vertical tissue slabs. At the microscopical stage it might not always be possible to identify the vertical axis within the field of vision. In this case low magnification micrographs of the section, including the horizontal plane normally solve this problem. For light microscopy a very useful device has been constructed, enabling rotation of the histological tissue specimen to align the vertical axis, see Fig. 9. Requirement 3: Relative to the common horizontal plane, the vertical sections must have random positions and random (i.e. isotropic) orientations. This is the first step in the sampling scheme
388

Test Systems for Vertical Sections Unbiased stereological estimates relying on point counting can freely be obtained from vertical sections, a fact, which enables the pathologist to make a number of stereological estimates from the same set of vertical sections. However, stereological estimates obtained by the use of test line systems must take into consideration the "vertical setting" of the sampling scheme described above. This lead to: Requirement 4: On the verticalsection, a test line is given a weight proportional to the sine of the angle between the test line and the vertical direction. This requirement is needed to make the test lines IUR lines in three-dimensional space, when used on a set of uniform random positioned vertical sections. The mathematical outline of this requirement has been decribed by Baddeley et al. 1986. Briefly, the probability of finding a IUR line at a given angle away from the vertical axis is proportional to the sine of this angle, (which increases from O at 0 to 1 at 90 away from the vertical axis). Fortunately, test systems based on these principles are rather easy to use. The weighting of test lines can most easily be obtained by construction of a test system in which test lines at a given angle to the vertical axis have a length proportional to the sine of this angle. Such test systems are needed for the unbiased estimation of surface area and their composition and application are described below, illustrated by an example. Estimation of Surface Area from Vertical Sections When used on vertical sections, test line systems, which fulfil requirement 4, mentioned above, are IUR in space. There are many ways to create correct test lines, see Baddeley et al. 1986, but for the practical purpose of surface estimation on vertical sections cycloids are the most suitable,

RECENT STEREOLOGICAL DEVELOPMENTS FOR PATHOLOGY

since only the total intersection number is then counted, and no numerical weighting is required, see Fig. 10. Test systems based on cycloids are made by computer assisted drawing and are most efficient when the test curves are "staggered" as shown in Fig. 11, thus avoiding unnecessary stastistical variation, e.g. caused by periodic effects in the biological structure under study, see Fig. 12. Surface area estimated on vertical sections, according to the principles outlined here, is free of any assumptions about shape. A very instructive example is given in the paper by Baddeley et al. 1986, illustrating both the production of vertical sections and the estimation by a cycloid test line system. However, the method has already been used within a range of biological settings. Chondrocytes of the epiphysal cartilage plate are anisotropically arranged, but vertical sections has solved the problem of unbiased estimation of cell surface area (Cruz-Orive & Hunziker 1986). Examples, where vertical sections have been generated for practical reasons, have also been carried out. For instance, estimation of capillary surface in the visual cortex has been done on vertical sections, and it is still possible to carry out other stereological estimates on the same tissue sections, since the cortical structure and orientation have been preserved, see Brndgaard & Gundersen 1986. A recent investigation includes the estimation of lung surface area on vertical sections, a practical solution in this case, where IUR sections are hard to make, Michel & Cruz-Orive 1988.

Fig. 11. a. Simple test system for estimation of surface

based on cycloids. The left edge of the frame must be aligned with the vertical axis of the section under study. SV is estimated by counting the number of intersections with the cycloid arcs, Vv of the reference space by point counting using the test points, whereas the straight lines are only introduced to make the following of the test system across the section easier. b. "Staggered" cycloid test system for improvement of statistical efficiency in layered structures. Produced by and available from Cruz-Orive, Anat. Inst., Berne University, Schweitz.

STAR VOLUME, A MEASURE OF SIZE OF VERY COMPLEX STRUCTURES

The star volume is defined as the mean volume of all parts of an object which can be seen unobscured in all directions from a particular point, see Fig. 13 (Serra 1982, Gundersen & Jensen 1985). The averaging in the above mean is over all points inside the structure. The star volume may be defined for any type of structure including cavities like bone marrow space and networks like the trabecular system. It owes its name to the fact that the structural characteristic it reports is clearly a volume and in very complex structures like bone it is almost the only way to obtain a measure of size with a strict mathematical definition. Knowledge of changes in the architecture of the human skeleton is important because the strength

of bone depends not only on the amount of bone present but also on the bone structure. Bone mass declines with age in both sexes due to a negative bone balance, that is, more bone is resorbed than formed. This loss of bone is not only due to a simple thinning of trabecular bone but there seems also to be a loss of bone caused by perforation or resorption of the trabeculae (Parfitt 1984). This leads to a discontinuity in the trabucular network and thereby to a decrease in bone strength.
389

GUNDERSEN et al.

Fig. 12. Vertical section of human skin projected onto a staggered cycloid test system for estimating surface area
of the dermoepidermal interface. The vertical axis is normal to the macroscopical cutaneous surface. Intersections between the dermoepidermal junction and cycloid arcs are counted. Due to the layering of the biological structure, staggered cycloids make the estimation more efficient. Points falling in the reference space (either epidermis or corium in this example) are also counted. (H + E, original magnification: 40X).

Conventional bone histomorphometry has hitherto not been able to demonstrate this pathogenesis - loss of trabecular bone components - due to a lack of unbiased stereological parameters. Such parameters can, however, now be obtained by

using vertical sections (Vesterby et al. 1987). One of the most informative stereological parameters in bone appears to be the star volume. The star volume of the marrow space and the trabeculae is very easily estimated: Whenever a

Fig. 13. Two-dimensional schematic drawing of the

star volume of the trabeculae and the marrow space. In 3 arbitrary units of volume, u , the marrow space star 3 volume at A is 109 u with a coefficient of variation (CV) 3 3 of 1.24, at B: 35 u (CV: 0.32), at C: 44 u (CV: 0.65). The 3 trabecular star volume at D is 13 u (CV: 0.81), at E: 16 u3 (CV: 1.5O) and at F: 3 u3 (CV: 1.79). The estimate of star volume varies very much, mainly as a function of orientation. The mean star volume of a structure is 3 times the average l 03 for many random points and many random directions chosen uniformly among all three-dimensional orientations.

388

RECENT STEREOLOGICAL DEVELOPMENTS FOR PATHOLOGY

Fig. 14. Sketch of a lumbar vertebra, indicating the sampling procedure for selection of two random vertical section planes, perpendicular to each other. The whole vertebra is positioned randomly with the inferior surface on a piece of paper which contains horizontal and longitudinal parallel lines spaced by 5 mm. The center of the vertera is marked and a rotation of the specimen is performed. The first random section plane is given by the longitudinal lines to the right of the center of the vertebra. (A1). The other perpendicular section plane is given by the horizontal lines, starting with the first one to the left of the center of the vertebra (B1). In this way approximately eight sections are obtained per vertebra. The section planes are parallel to the "cylindrical" axis of the vertebra which is also the vertical axis.

Fig. 15. Test system for estimating star volume from vertical sections consisting of a circular frame for selecting directions and an inner system of lines and points. The direction of the outer frame axis is indicated by the arrow at the top, the frame is fixed with this direction parallel to the vertical direction. The construction of the intervals is illustrated in Fig. 16. The frame is divided into 97 directions; for the first field a random number is selected (e.g. 3) and the transparent, inner grid which contains parallel lines and test points is rotated into that direction. The next direction in the same or the following fields is selected by adding a predetermined number - for instance 20 - to the number of the previous selected direction (3 + 20 = 23). When the sum exceeds 97 (e.g. l13) it is reduced by 97 (113 - 97 = 16) and the remainder (16) is the new number and thereby the new direction.

random point on a random section hits an object one measures the length l0 of the unbroken linear intercept in the structure through the point in a three-dimensionally isotropic direction. Each intercept length is then raised to the third power and the star volumen, v* is estimated unbiasedly by
* 3 v = I0 3

The statistical variation of the estimate is often rather large, and in order to obtain reasonably narrow confidence limits it is likely to be necessary to measure 100 to 200 intercepts in as many different directions as possible - in three-dimensignal space. The unbiasedness of the estimator is most critically dependent upon the three-dimensignal isotropy of the chosen direction in which the

intercept is measured. Finally, the estimate is only biologically meaningful if sampling is uniform over a suitably defined reference space, e.g. a complete bone. The sampling scheme developed for vertebral corpora aims specifically at these problems. The practical sampling procedure is as follows: In order to produce random section planes parallel to the vertical axis of the vertebrae each specimen is placed with the inferior surface on a sampling grid with horizontal and longitudinal parallel lines spaced 5 mm, see Fig. 14. The vertebra is then randomly positioned and randomly rotated to find the first plane of sectioning defined by the lines of the sampling grid, the center, and the cylindrical axis of the vertebra. In this randomly selected orientation three to five 5-mm-slices are cut con391

GUNDERSEN et al.

Fig. 16. Construction of non-equidistant, sine-weighted orientations for selecting three-dimensionally isotropic directions on vertical sections. On the radius shown to the left of the quarter circle a number of equidistant intervals are marked off. It is most efficient to use a prime number. The end-point of each interval is transferred horizontally to the circular arc. Directions from the centre of the circle through these points are sine-weighted, i.e. their density is proportional to the sine of the angle from the vertical axis to the direction itself. See also Figs. 6 and 15.

in one direction at a time to provide a systematic pattern of touching fields. On an average, nine fields per section are used for estimating the star volumen of trabeculae and two fields per section for the marrow space. The test system consists of a frame for systematic selection of random threedimensional isotropic directions on vertical sections and of a set of points and parallel lines on a separate, transparent sheet, see Figs. 15 and 16. The axis of the frame is kept parallel to the vertical axis of the sections. The intercept lengths are measured using a ruler separated equidistantly into 20 classes. On an average, 172 measurements for marrow space and 155 for trabeculae are performed per specimen, see Fig. 17. The material for the preliminary study reported here consisted of the first lumbar vertebra which was obtained at autopsy from two females and seven men with a mean age of 57 ages (range 26 to 79 years) without malignant diseases or metbolic bone diseases. The star volume of the marrow space increased from 11.5 mm3 in the youngest to

secutively on a sawing machine. In a plane perpendicular to the first randomly selected plane three to four similar slices are cut. That is, from each vertebra an average of eight 5-mm-thick slices are cut in two mutually perpendicular planes and all slices are parallel to the cylindrical axis of the vertebra, which also is the vertical axis. Bones slices are embedded, undecalcified, in methylmetacrylate. From each direction two slices are selected systematically randomly and from each of these four slices per vertebra one 8-mthick section is cut on a heavy duty microtome (Jung, model K) and stained with GoldnerTrichrome. Using a recently developed special projection microscope (Olympus, BHS) the image of the sections is projected onto a test system placed on the table at a final magnification of 16X for the estimation of marrow space star volume and of 65X for the estimation of the star volume of trabecular bone. Starting at a random position outside the section the microscope stage is moved
390

Fig. 17. Two-dimensional schematic drawing of marrow space and trabeculae of a lumbar vertebra at a magnification of approximately 16X, at which magnification the star volumen of the marrow space is measured. All 12 encircled points are used for sampling in the marrow space. In the shown example, 10 sampling points fall on the marrow space and through all these the linear intercept length - in the direction of the lines - of the marrow space are measured and raised to the third power ( l ). If two sampling points on a line fall on the "same" marrow space the intercept is measured twice, see also Vesterby et al. 1988.
3 0

RECENT STEREOLOGICAL DEVELOPMENTS FOR PATHOLOGY

61.5 mm3 in the oldest in the group; marrow space star volume was related to age with a coeffiicient of correlation of r = 0.81, 2p = 0.014 (Vesterby et al. 1988). The slope of the regression line indicated that the marrow space star volume increases 1 to 2 mm3 per year in normal man. It is quite remarkable that the trabecula star volume does not change with age: it was 0.17 mm3 in the youngest and 0.23 mm3 in the oldest with no indication of any relationship to age. All told, the only way to get such a pronounced increase in the size of the marrow space star volume is by removing or perforating bone trabeculae. The whole structure thereby becomes more and more spongy with age, but the trabeculae remaining in old people have sizes similar to those in young people. Therefore, the star volume looks a very promising parameter for elucidating the pathogenesis of architectural bone changes; it is the only parameter which gives a direct and unbiased estimate of a well defined size of the cavities in the marrow space. However, to confirm its superiority as a tool in this type of bone research it has to be applied to larger groups of normal persons and to osteopenic diseased patients. In addition to the above example, star volume is likely also to be very useful for obtaining a well defined size parameter for very irregular cells, a problem with many aspects which are dealt with in the next part of this review.

REFERENCES
Baddeley, A. J., Gundersen, H. J. G. & Cruz-Orive, L. M.: Estimation of surface area from vertical sections. J. Microsc. 142: 259-276, 1986. Brndgaard, H. & Gundersen, H. J. G.: The impact of recent stereological advances on quantitative studies of the nervous system. J. Neurosci. Meth. 18: 39-78, 1986. Cruz-Orive, L. M. & Hunziker, E. B.: Stereology for anisotropic cells: Application to growth cartilage. J. Microsc. 143: 47-80, 1986. Gundersen, H. J. G.: Notes on the estimation of the numerical density of arbitrary profiles: The edge effect. J. Microsc. 111: 2 19-223, 1977. Gundersen, H. J. G.: Estimators of the number of objects per area unbiased by edge effects. Microsc. Acta. 81: 107-117, 1978. Gundersen, H. J. G.. Stereologi, eller hvordan tal for rumlig form og indhold opnas ved iagttagelse af strukturer pa snitplaner. Bibl. Lger. 172: 43-61, 1980.

Gundersen, H. J. G. & Jensen, E. B.: Stereological estimation of the volume-weighted mean volume of arbitrary particles observed on random sections. J. Microsc. 138: 127-142, 1985. Gundersen, H. J. G.: Stereology and sampling of biological surfaces. In: Echlin, P. (Ed.): The Analysis of Organic and Biological Surfaces. J. Wiley & Sons, New York 1985, chapter 19, pp. 477-506. Gundersen, H. J. G.: Stereology of arbitrary particles. A review of unbiased number and size estimators and the presentation of some new ones, in memory of Wiliam R. Thompson. J. Microsc. 143. 3-45, 1986. Gundersen, H. J. G. & Jensen, E. B.: The efficiency of systematic sampling in stereology and its prediction. J. Microsc. 147: 229-263, 1987. Gundersen, H. J. G., Boysen, M. & Reith, A.: Comparison of semiautomatic digitizer-tablet and simple point counting performance in morphometry. Virschows Archiv. 37: 3 17-325, 1981. Gundersen, H. J. G., Gtzsche, O. & sterby, R. Sampling efficiency in morphometry simplified. Metab. Bone. Dis. and Rel. Res. 2: 443-448, 1980. Gundersen, H. J. G. & sterby, R.: Optimizing sampling efficiency of stereological studies in biology or "Do more less well"! J. Microsc. 121: 65-73, 1981. Jensen, E. B. & Gundersen, H. J. G.: Stereological ratio estimation based on counts from integral test systems. J. Microsc. 125: 161-166, 1982. Jensen, E. B., Gundersen, H. J. G. & sterby, R.: Reconstruction of membrane thickness distribution from orthogonal intercept lengths in a plane section. Scand. J. Statist. 6: 182-183, 1979. Mattieldt, T., Mall, G., Von Herbay, A. & Moller, P.: Stereological investigation of anisotropic structures with the orientator. Acta Stereol. 8: 671-676, 1989. Mattieldt, T., Mobius, H.-J. & Mall, G.: Orthogonal triplet probes: An efficient method for unbiased estimation of length and surface of objects with unknown orientation in space. J. of Microsc. 139: 279-289, 1985. Michel, R. P. & Cruz-Orive, L. M.: Application of the Cavalieri principle and vertical sections method to lung: Estimation of volume and pleural surface area. J. Microsc. 150. 117-136, 1988. Nyenguard, J. R., Bendtsen, T. B. & Gundersen, H. J. G.: Stereological estimation of the number of capillaries, exemplified by the renal glomerulus. APMIS Supl. 96: 92-99, 1988. Parfitt, A. M.: Age related structural changes in trabecular and cortical bone: Cellular mechanisms and biomechanical consequences. Calcif. Tissue Int. 36: 123-128, 1984. Pakkenberg, B.: Postmortem study of chronic schizophrenic brains. Brit. J. Psychiat. 151. 744-752, 1987. Regeur, L. & Pakkenberg, B.: Optimazing sampling designs for volume measurement of components of human brain using a stereological method. J. Microsc. 155: 113-121, 1988. Seyer-Hansen, K., Hansen, J. & Gundersen, H. J. G.:

391

GUNDERSEN et al.

Renal hypertrophy in experimental diabetes: A more phometric study. Diabetologia. 18: 501-505, 1980. Serra, J.: Image Analysis and Mathematical Morphology. Academic Press 1982, London. Vesterby, A., Kragstrup, J., Gundersen, H. J. G. & Melsen, F.: Unbiased stereological estimation of surface density in bone using vertical sections. Bone 8: 13-17, 1987. Vesterby, A., Gundersen, H. J. G. & Melsen, F.: Unbiased stereological estimation of osteoid and resorption fractional surfaces in trabecular bone using "vertical sections". Sampling efficiency and biological variation. Bone. 8: 333-337, 1988. Weibel, E. R.: Stereological Methods. Practical Methods for Biological Morphometry, Vol. 1. Academic Press 1980, New York. sterby, R. & Gundersen, H. J. G.: Sampling problems in the kidney. In: Miles, R. E. & Serra, J. (Eds.): Geometrical Probability and Biological Structures:

Buffon's 200th Anniversary. Lecture Notes in Biomathematics, No. 23, Springer-Verlag, Berlin, 1978. sterby, R. & Gundersen, H. J. G.: Fast accumulation of basement membrane material and the rate of morephological changes in acute experimental diabetic glomerular hypertrophy. Diabetologia. 18: 493-501, 1980. sterby, R. & Gundersen, H. J. G.: Stereological estimation of capillary length exemplified by changes in renal glomeruli in experimental diabetes. In: Reith, A. & Mayhew, T. M. (Eds.): Stereology and morephometry in electron microscopy, problems and solutions. Hemisphere Publishing Corporation, New York 1988, pp. 1 13-122. The list of references is updated in this 2nd printing in 1990, but the references in the text are unchanged.

390

APMIS 96. 857-881, 1988

The new stereological tools: Disector, fractionator, nucleator and point sampled interepts and their use in pathological research and diagnosis
Review article H. J. G. GUNDERSEN, P. BAGGER1, T. F. BENDTSEN, S. M. EVANS2, L. KORBO3, N. MARCUSSEN, A. MLLER3, K. NIELSEN4, J. R. NYENGAARD, B. PAKKENBERG3, F. B. SRENSEN, A. VESTERBY5 and M. J. WEST Stereological Research Laboratory, University Institute of Pathology, 2nd University Clinic of Internal Medicine, Institute of Experimental Clinical Research, University of Arhus, 1Laboratory of Reproductive Biology II, University of Copenhagen, Rigshospitalet, 2Department of Psychiatry, University of Liverpool, England, 3 Neurological Research Laboratory, Hvidovre Hospital, 4Institute of Pathology, University of Copenhagen, Rigshospitalet, and 5University Institute of Pathology, Arhus Amtssygehus, Denmark Gundersen, H. J. G., Bagger, P., Bendtsen, T. F., Evans, S. M., Korbo, L., Marcussen, N., Mller, A., Nielsen, K., Nyengaard, J. R., Pakkenberg, B., Srensen, F. B., Vesterby, A. & West. M. J. The new stereological tools: Disector, fractionator, nucleator and point sampled intercepts and their use in pathological research and diagnosis. APMIS 96: 857-881, 1988. The new stereological methods for correct and efficient sampling and sizing of cells and other particles are reviewed. There is a hierarchy of methods starting from the simplest where even the microscopic magnification may be unknown to the most complex where typically both section thickness and the magnification must be known. Optical sections in suitably modified microscopes can be used to improve the ease and speed with which even the most demanding of these methods are performed. The methods are illustrated by practical examples of applications to a wide range of histological entities including synapses, neurons and cancer cells, glomerular corpuscles and ovarian follicles. Key words: Disector; fractionator; nucleator; selector. H. J. G. Gundersen, Stereological Research Laboratory, University Bartholin Bygning, DK-8000 rhus C, Denmark. This is the second part of a review on New Stereological Tools by Gundersen & Coworkers. The first part appeared in APMIS, vol. 96, Number 5, May 1988, pp. 379-394.

COUNTING AND MEASURING IN THE REAL 3-DIMENSIONAL WORLD

The first part of this review (Gundersen et al. APMIS 96: 379-394, 1988, henceforth referred to as APMIS-1) dealt with a number of basic stereological principles for 2-dimensional quantitation and their use as tools for obtaining efficient and unbiased quantitative estimates of 3-dimensional structures. The methods required only a single,

uniformly positioned thin section. With one exception, the methods in this second part of the review require two parallel sections which are often, but not always, of known thickness and specified orientation. The two sections need not be physical sections, i.e. physically separate, since optical sections suffice in almost all instances and are much more efficient, as decribed in detail below. The methods described here represent a major
857

GUNDERSEN et al.

disector (Sterio 1984), it is necessary to describe a principle for measuring n-dimensional content and its spatial distribution. It was not described in APMIS-l because it is as meaningless and useless in two dimensions as it is strong and versatile in three.

THE NUCLEATOR
We all learned in school that the area a of a circle is r2, where the radius r is defined as the (constant) distance from the centre to the boundary. However, the same relation holds if we measure the distance l from any fixed point in the circle to the boundary, but only if the direction in which we measure is isotropic, see Fig. 1. Since the distance l now varies, the unbiased estimator of the area is a= l 2 , i.e. if we measure in more than one isotropic direction the distances are squared before averaging. Note that a r2 is an identity because r is constant, whereas a= l 2 is an estimator with a certain coeffiicient of error, CE. Unbiasedness means, as always, that if we measure l in sufficiently many isotropic directions, the mean value of l 2 comes arbitrarily close to the true area (and the CE goes to zero). This old and well known relationship is true for any shape, however, and it does not require the fixed point to be in a specific position. The point may even be outside the profile, see examples in Gundersen 1988. In 2-dimensional space this relationship is more of a curiosity (but how was it the planimeter worked?) than a working formula for estimating profile area, point counting is always likely to be more effiicient, because we can see and sample the complete profile. In three dimensions it is quite another story. Ordinarily, we can only observe 2-dimensional sections and despite the notable effiiciency of the Cavalieri-estimator of volume (APMIS-l) this estimator does require exhaustive sectioning of the object. The above estimation principle of the nucleator works, however, in any n-dimensional space: from an arbitrary, fixed point measure the distance l to the boundary in any isotropic direction and content = c l
n

Fig. 1. The area of a circle may be estimated in two ways: a - r2 and a= l 2 , where the first is in fact just a special case of the latter. If the measured distance from the arbitrary point to the boundary varies with orientation one must measure in a random direction (uniform in the interval from 0 to 360). The estimator is independent both of the shape of the profile and of the position of the point. For more than one intercept from the point one measures all squared distances to the boundaries and adds the ones which terminate over the profile and subtracts those which terminate on the outside of the profile. It is always an advantage to measure in both directions from the point. If the point is outside there are directions in which the measure is zero; they must also be taken into acount in the mean, as shown.

breakthrough in Stereology because they are truly 3-dimensional and because they are the only stereological methods whereby particle number and sizes can be estimated in an unbiased manner. Before describing the principle on which this new and rapid development of Stereology is based, the
858

where for n = 1, 2, 3,... "content" is length, area, volume,... and c = 2, , 43 ,.. (try it for n = 1!). For a 3-dimensional object, this means that the

THE NEW STEREOLOGICAL TOOLS

TABLE 1. The four stereolocical Drobes and the ceometric characteristic they estimate in 3-dimensional objects

Stereological Probe Name Measure Point number Line length Plane area Disector volume

Dimension 0 1 2 3

Geometric characteristic of 3-dimensional object Name Measure Dimension Volume volume 3 Surface area 2 Curve length 1 Cardinality number 0

A list of the four basic stereological probes with indication of their measure (how to express the amount of the probe) and dimension. To the right the geometric characteristic of 3-dimensional objects to which the probe is "sensitive" (and therefore can estimate) is shown for each probe. Note the ambiguity of the word "volume". A given probe of dimension k is a direct estimator of a d-dimensional characteristic only if d = n - k where n is the dimension of the space embedding the object: l-dimensional lines are therefore the pnncipal probes for estimating (3 - 1) = 2-dimensional surface area in 3-dimensional space. These are the principal or "primitive" relations where the estimation is performed by counting only, other relations exist as well. In fact, any k-dimensional probe can be used as an estimator of all geometric characteristics with dimension d n - k, but not by just counting: the fractional length of test lines, LL , is an estimator of both areal fraction AA and volume fraction VV but one must measure the lengths of the intercepts. Note that whenever d= n or k = n one can perform the estimation ignoring the orientation distribution of both the probe and the object, in all other cases one has to know the orientation distribution of either object or probe. In practice this means that one must use IUR sections for estimating length or height with 2-dimensional planar counting frames and IUR lines on vertical or IUR sections for surfaces and for all direct size or distributional estimators based on the nucleator.

volume of the object can be estimated on just one section through the fixed point. The section must be either isotropic or fulfill the requirements for a "vertical" section to enable us to measure in isotropic directions in 3-dimensional space. Note that in practice the point must be unique and recognizable for it to remain in a fixed position independent of the direction of the section. These conditions can all be met in mononucleated cells, see Gundersen 1988, and, even more efficiently, in cells with just one nucleolus. Other structures may also be analyzed this way, cf. the follicle in Fig. 15 below. For many more details, proofs and a large number of closely related estimators, see Gundersen 1988. The nucleator principle: measurements carried out isotropically with respect to a fixed point, is very versatile. It is possible to estimate the surface of arbitrarily shaped, nucleated cells, Gundersen & Jensen l987b. In addition, a whole range of other properties may be estimated as well, see Jensen et al. 1988, the example below illustrating courting with the nucleator? and Evans & Gundersen 1989a and Fig. 16 below illustates how it can be used to estimate the spatial distributions of all these quantities. However, before going into more details it is first necessary to describe

how to sample particles uniformly, of necessity the first step in analyzing them.

THE DISECTOR: HOW TO SAMPLE OBJECTS UNIFORMLY

Humans have five senses like most living creatures: Smell, sight, touch, taste, and hearing. We do not generally get confused because a given object (a dead fish, maybe) is both visible, stinking and feels slimy at the same time. These are different qualities (modalities) of the same object. As opposed to most lower species, stereologists have at least four extra senses or probes which are sensitive towards and very specific for four geometric modalities of 3-dimensional objects: volume, surface, length, and number, see Table 1. The real problem used to be that no probe can estimate a characteristic of dimension less than the (n - k) dimensions of its principal target. Therefore none of the first three probes listed in Table 1 (i.e. those reviewed in APMIS-l) can estimate 0-dimensional number, only the 3-dimensional disector "feels" number without regard to volume, shape or height of the isolated objects.

859

GUNDERSEN et al.

Fig. 2. The disector. Two parallel section planes a known distance h apart with an unbiased counting frame of area a(frame) on the sampling or reference plane. Complete transects (one or more profiles in the same particle) are sampled if they are partly or totally inside the frame provided they do not in any way intersect the fully drawn exclusion edges or their extension. There are Q = 4 such transects sampled in the Figure. Of these four, two- are intersected by the upper look-up plane and are not counted. The number of particles in the probe is the remaining Q = 2.

Like the unbiased 2-dimensional sampling and counting frame show in Fig. 2 and described in APMIS-1, the disector is a probe which samples isolated objects or particles with a uniform probability in 3-dimensional space, irrespective of their size and shape. The disector is shown in Fig. 2. The section pair must have uniform position in the reference space, the space which contains the particles, and generally the sampling frame must have a (systematic) random position in the reference section. The precise rule for sampling particles is given in the legend to Fig. 2, see also the original description (Sterio 1984) and the review of the disector and many other estimators based on this principle in Gundersen 1986. For an uncovering of the almost 100-year-long history of the disector-principle, see Bendtsen & Nyengaard 1988. "Sampling particles with uniform probability" means that a priori all particles have the same chance of being sampled. The set of particles actually sampled therefore constitutes a "fair" or "representative" or uniform sample of all particles. In practical terms, this means that the mean characteristic of the sample is an unbiased estimate of the same mean characteristic of all particles in the reference space. This simple and immutable fact is the basis for many of the applica-

tions of the disector. The height h of the disector (usually the thickness of one of the two adjacent sections) need not be known - sampling is uniform anyhow. A disector-sample is the only stereological sample of particles which has this extremely importent feature of uniformity. Specifically, a sample of cells taken because they were hit by a single section is not a uniform sample. Such a sample is height-weighted: a cell which is twice as high as another has twice the chance of being in the sample. This is worth emphasizing: the cells (parti-cles) seen in a single section is a biased, non-representative sample of all cells. The simplest use of a sample is to count the Qelements in it. Knowing the magnification and the height h, the volume v(dis) of the disector-probe is known: v(dis) = a(frame) h and a straightforward estimator of the total number of particles in a specimen of volume V(ref) is:

N(particles) =

Q v ( dis )

V ( ref )

where the summation is over a number of systematically sampled disectors in the reference space. In biology, the tissue surrounding the particles to be counted or sampled is almost always transparent and the section thickness may therefore be larger than the minimal diameter of the particles to be

860

THE NEW STEREOLOGICAL TOOLS

counted (Sterio 1984). In order to identify and distinguish different particles correctly one should rarely use a section thickness exceeding 1/4 to 1/3 of the height of the particles, which in practice is the more important constraint on section thickness. If the particles are sparsely distributed one may use slightly thicker sections, whereas for densely packed particles one may have to use very thin sections - or rather optical disectors if this is at all possible. The two situations are illustrated in Figs. 4 and 15 which show neurons and follicle granulosa cells, respectively. (Some special problems arise when the absence of something is counted, e.g. holes in perforated synapses, see Calverley & Jones 1987) In many cases one may take advantage of the freedom to use a particularly "good" direction for identifying the particles to be counted, see for instance Mulvany et al. 1985, who counted smooth muscle cell nuclei in arterioles on sections parallel to the vessel axis and thereby perpendicular to the long, curved cell nuclei. Sectioned this way it was much easier to identify the nuclei. Undoubtedly the most common question asked by investigators starting stereological projects is: "How many must I count?", i.e. points, intersections, profiles or particles. The answer is simple: "never count more than 100 to 200 (and that only in extreme cases) in one biological unit (often an organ)!". For a general consideration of this, see the discussion and review in Gundersen 1986, Gundersen & Jensen 1987a, Michel & Cruz-Orive 1988, and the detailed example and calculation of the pertinent CEs in Pakkenberg & Gundersen 1988. To our knowledge, a courter-example of this rule based on a statistical analysis of sampling designs has never been published.

Fig. 3. The modified Olympus BH-2 microscope has an electronic microcator (Heidenhain, VRZ 401 ) attached to the side (M), and a projection arm. The light source is a 100 Watt halogen light bulb which is housed in a modified lamp housing to maximize its illuminating power. A set of motors (not shown here) is usually attached to the specimen stage for stepwise, predetermined movements in the x- and y-direction to aid in the systematic sampling of fields of view. There is a special rotating stage fitted (see Fig. 9 in APMIS-1) so that specimens can be rotated through 360 independently of the x,y-movement of the stage (When in use, the projection is onto test-systems on the table surface).

THE OPTICAL DISECTOR

The idea of the optical disector (Fig. 2.3 in Gundersen 1986) is to start by making only one, relatively thick section, ~ 25 m, say, and then make the two or more parallel section planes for the disector as thin optical sections inside the thick one by moving the plane of focus up or down. Optical sectioning greatly facilitates the application of most of the new stereological techniques described in this review. Optical sectioning will be at its best when used on a confocal microscope, see e.g. Petran et al. 1968 and Howard et al. 1985, a

wonderful invention which still does not work in transmission, only in reflectance or fluorescence illumination mode. Almost all the methods illusbated below have therefore been carried out in thick plastic sections on a conventional light microscope which in various ways has been modified for stereological use, see Fig. 3. The two main modifications are an electronic microcator which has been fitted to the microscope to measure the movements of the stage in the z-axis and that the microscope is fitted with a projection arm. High numerical aperture oil immersion lenses, with a matched numerical aperture condenser give the smallest possible focal depth. Only oil immersion lenses should be used for optical sectioning because the movement of the specimen stage, which the microcator measures, then equals the movement of the plane of focus (provided the refractive indices of the embedding medium and of the immersion oil are equal). Using glycolmethacrylate and, for example, a standard Giemsa staining, the thickness of the physical section can be varied from 0.5 m to more than 100 m without problems. Although disec861

GUNDERSEN et al.

Fig. 4. Two optical sections 2 m apart in the middle of a Giemsa stained glycolmethacrylate section of ~ 25 m thickness. In an appropriately (systematically) sampled microscopic field, one focuses down ~ 5 m to avoid distortions and unevenness of the section surface. All neuron nuclei seen in focus in this look-up plane to the right are disregared. Then one focuses through a distance of e.g. 10 m in 2 m steps counting any neuron nucleus which comes into focus, Including those in the last optical section. A neuron is counted or sampled when its clearly focussed nucleolus is counted in the disector sampling frame as shown to the left. A non-negligible fraction of small neurons in human cortex do not have a well-defined neucleolus, potentially quite a problem when in the old days neuron nucleoli were counted in one section only (?). By contrast, in the optical disector one focuses through the nucleus and observes the lack of a nucleolus and instead simply counts these neurons when their sharpest nuclear profile falls within the disector counting frame. The physical section should be sufficently thick to allow a guard volume of a few m thickness at the bottom of the disector for resolving ambiguities in identification and to avoid the problems of lost caps. The frame size is 67 m by 137 m. The ruler used for classifying intercepts to either side of the nucleolus or nucleus centre (for nucleator-estimates of neuron volumes) is shown below at ~ 3 times normal size. It is constructed like the one shown in Fig. 6.

862

THE NEW STEREOLOGICAL TOOLS

tors can be made from physical sections there are important advantages associated with optical sectioning. The most time consuming step when using physically separate sections as disectors, that of aligning the two sections so that the two fields of view in two identical microscopes coincide, is eliminated, see Fig. 4. This simple fact reduces the time spent on the counting or sampling of cells to roughly a fifth, making it possible to estimate for example the total number of neurons in a human brain cortex with a CE of ~ 5 per cent in less than half a day; and the estimate is then even made for separate cortical areas (Brndguard et al. 1989). The extra information obtained from working in three dimensions also makes the identification of individual structures much easier. In fact, very densely packed cells like granule cells in cerebellar cortex and granulosa cells in follicles, see Fig. 15, are almost impossible to count in physical disectors. Moreover, since the microcator is fitted to the side of the microscope, the height of a disector is measured directly, (the electronic microcator has an accuracy of more than 0.5 m) thereby avoiding the problem of elaborate calibration of the microtome or of measuring the thickness of the physical section. Cells with more than one nucleus or more than one nucleolus used to present a rather awkward problem when applying the disector or the nucleator principle, since both techniques are only effortless to apply when the particle contains something unique inside. So it is another advantage of optical disectors that given sufficient section thickness, one can sample nuclei or nucleoli in the middle 5 to 10 m of the section and then use the rest of the thickness to ensure that sampling was unique and thereby uniform by looking through the sampled cells and determining the number of nuclei or nucleoli in them, see Vesterby 1989 for an example of polynucleated cells.

POINT-SAMPLING OF NUCLEAR INTERCEPTS IN CANCER GRADING

The principal sensitivity of a probe to a certain geometric characteristic, listed in Table 1, is a feature which is used directly in the estimation of these characteristics. The sensitivity of a probe is also used in sampling schemes, most directly in the sampling of particles. As mentioned above, one

section plane samples particles with a chance proportional to their linear dimension ("length" or height), a sampling scheme only of interest here because it produces a non-uniform sample of a type at which almost all microscopists look every day. Sampling particles with points has been used in geology and materials sciences for many years. This type of sampling was revived for biological use a few years ago (Gundersen & Jensen 1983 and 1985). Random points thrown on a randomly positioned section hit a particle with a probability which is directly proportional to the volume of the particle. Sampling therefore only cells hit by random points, one studies a volume-weighted sample of cells or a sample of cells from the volumeweighted distibution. There are two good reasons for using this special sampling technique in certain areas of research and diagnosis. In biology there are many instances in which one might expect the largest particles to carry the most information about a certain change. Studying a sample of cells which contains a larger fraction of big nuclei than a uniform sample might therefore lead to a greater sensitivity in detecting early or relatively small changes (Gundersen 1986). Another reason for using point sampling as opposed to uniform sampling with the disector, is that it requires only one section. An attractive feature of this scheme is that the point-sampled particles in the section can then be studied by the nucleator sizing-principle on the same section. When, in the description of the nucleator, it was stressed that the point must be unique and recognizable, the reason was mainly to ensure correct uniform sampling with the disector. In an already sampled cell the nucleator sizing principle works for any point inside the cell, including a random sampling point. Since the point is random inside the particle, it is not effiicient to make two measurements in two opposite directions from the point, it is better to measure the length of the complete intercept through the point in a 3-dimensionlly isotropic direction (on IUR or "vertical" sections). The coeffiicient to be used for calculating the volume weighted mean particle volume vV is then 3 instead 4 of 3 . This special case of the general nucleator principle was described before the nucleator and has its own name, point-sampled linear intercepts, which also describes reasonably precisely what it involves. One of the most uniform features of many malignant cell populations is the change in their
863

GUNDERSEN et al.

nuclear morphology: the appearance of cells with nuclei which are quite large. Since the absolute size is difficult to judge with any precision, the variability of nuclear size between cells may often be the most telling feature. Tumours with very large cell nuclei or showing a pronounced variability of nuclear size will generally be deemed more malignant. It is well recognized among pathologists that the subjective assessment of these changes on a qualitative scale is not very reproducible. Various quantitative methods have therefore been used with varying success (e.g. Helander et al. 1984, and Ooms et al. 1985), but most of these have used only 2-dimensional quantities as a basis for making decisions. Such techniques are unlikely to be optimal for 3-dimensional changes in 3-dimensional structures. In routine histopathology many sections of tumours fulfill the criteria for uniform random point sampling and isotropically oriented test lines. Tissue chips from tumours, e.g. from bladder tumours and from prostatic tumours may be assumed to be approximately randomly oriented during the embedding procedure. This means that routine sections of such tumour specimens for all intents and purposes are IUR sections. One may then use a sampling probe of systematic points on lines in any, fixed direction (parallel to the edge of the table, for example). Using such a test system ~ 75 intercepts are classified on 5 to 10 fields of vision using a projection microscope such as that shown in Fig. 3. This classification takes less than 10 minutes per tumour using the ruler shown in Fig. 6. Knowing the magnification and the physical length of the ruler one calculates the mean, cubed intercept length, as described in detail in Brndguard & Gundersen 1986, and obtains the 3 estimate v v = 3 l 0 . The CE depends on tumour type and grade but it is mostly in the range 0.1 to 0.3, a satisfactory level for routine use in single patients and for scientific purposes. It should be noted that this estimate of the mean nuclear volume is both objective and very reproducible (Nielsen et al. 1988a). A good correlation between the mean nuclear volume in bladder tumours and the prognosis is present. With arbitrary cut-off points of ~ 300 m3, 300 to 500 m3, and > 500 m3 80 patients were divided into three subgroups with 35, 26, and 19 patients, respectively. Five years after the diagnosis, l , 7, and 18 in the respective subgroups had died from their cancer (Nielsen et al. 1986). In a
864

Fig. 5. The recurrence pattern in a ten-year follow-up period in 35 patients with urinary bladder papilloma as a function of the volume-weighted mean nuclear volume in the first biopsy. The median vv of the group is 165 m3. The areas of the black rectangles are proportional to the frequency in each row. Modified from Nielsen et al 1988b.

time where resources for health care seem to becoming more scarce, it may also be of value for the distribution of these resources that another study showed a very strong relation between the mean nuclear volume in a patients first bladder papilloma and the probability that over the next 10 years the patient 1) would not develop a new tumour or 2) would develop one or more tumours which would not become invasive or 3) would develop invasive bladder cancer, see Fig. 5 and Nielsen et al. 1988b. Biologically, it is remarkable that although all patients had their initial tumour removed, the nuclear vv of the first tumour nevertheless was closely related to the malignant potential of the next tumour(s), making it possible to assign a different frequency of control visits to the clinic on the basis of the first biopsy. In the above example, the isotropic orientation of the section plane was easily made (see also the "Cucumber" Orientator by Mattieldt in APMIS-1) and was easy to use. There are, however, a number of tissues where too much information about the precise location of the samples is lost on IUR sections, as in tubular systems or surface epithelia like the example of a benign cutaneous naevus described here. The advantage of the vertical section design is that the tumour is sliced into vertical sections (systematically around the vertical axis for optimal sampling, see APMIS-1),

THE NEW STEREOLOGICAL TOOLS

Fig. 6. A vertical section of a benign cutaneous naevus projected onto the test system for vertical sections. The vertical axis has been aligned with the left hand edge of the orientation frame at a lower magnification. For each point inside the frame which hits a nucleus, the nuclear intercept through the point is measured in the direction dictated by the test-line. All points which fall inside the frame are considered (also if they hit a nucleus which is on the edge of the fame) whereas points falling outside the frame are ignored. The direction-number for the first field is a random number between 1 and 97, in the Figure 70 was obtained from a random number table, e.g. page 131 in Documenta Geigy, Scientific Tables, 7th ed., Diem & Lentner 1970. For each of the following fields of vision a new direction-number is obtained by adding 37 to the previous number or, if the sum is > 97, subracting 60. The next numbers after 70 are therefore 10, 47, 84, 24, etc. (Giemsa stain, original magnification: l500X). In the Figure a classified ruler for measurements of intercepts is also seen. The ruler is made with a scale of volume or length raised to the third power to obtain cubed intercept lengths directly. The widths of the first and the 15th class have a ratio of 10:1, i.e. on the volume scale the 15th class is ten times wider than the first, almost the opposite of the physical ruler seen here on an of necessity linear scale. Due to this construction, the ruler classes are differentiated and handy to use, for instance in sections of nuclear populations showing size-pleomorphy, see also Fig. 10 where an even more differentiated ruler was used.

perpendicular to the horizontal plane of reference which is the macroscopical epidermal surface. Vertical histological sections are projected onto a table at a final magnification of ~ 2000X using the projection microscope shown in Fig. 3. Nuclei are point-sampled and the point-sampled intercept lengths are measured according to the principle used for isotropic sections. The crucial step is the construction and application of a test-line system

which satisfies the requirement for three-dimensional isotropic test-line directions. In brief, testlines with a greater angle to the vertical axis must be selected with greater frequency than those close to the axis. For the practical implementation of such a system a "direction-finder" or orientation frame can be constructed, see Fig. 16 in APMIS-1, where the left-hand edge is always aligned with the chosen vertical axis. A transparent test-system
865

GUNDERSEN et al.

with points and lines is then superimposed on the orientation frame with a systematic random choice of direction-number, see Fig. 6. Any ruler may be used for the measurements of intercepts, but a handy ruler construction whereby one actually obtains cubed intercept lengths is also shown in Fig. 6. Detailed description of the ruler design and calculated examples of vv are found in Brndgaard & Gundersen (1986) and Srensen (1989c). Estimates of the volume-weighted mean nuclear volume from vertical sections of cutaneous melanocytic tumours (Howard 1986, Brungger & Cruz-Orive 1987, and Srensen 1989a & l989b) turn out to be closely related to the tumour type. More importantly, it is also highly correlated to the patients' five-year survival, even in a homogeneous group of patients with malignant melanomas all in stage 1 , see Srensen 1989a & l989b. Estimates of nuclear vv may be of especially great value in the search for objective malignancy grading parameters in various cancers. One reason why the estimator is a very sensitive parameter is that it combines information on both mean nuclear 2 size and variability of size, since VV = VN (1 + CV N (v)) where v N is the mean volume in the ordinary number distribution and CVN(V) is the coeffiicient of variation of particle volumes in the same distribution. In fact, by estimating nuclear volume in both the number distribution and the volume distribution nuclear pleomorphy may be defined quantitatively, another promissing prospect within the field of future objective malignancy grading (Srensen 1989c).

THE FRACTIONATOR
In the rest of the methods in this review particles will be considered as individuals of equal importance, i.e. mean volumes are estimated in the number distribution and total number is also a parameter of direct importance. All the methods therefore rely on the disector as the primary sampling device. In most of the examples the disector is used in a very direct way without knowing for instance the section thickness. The simplest of these sampling schemes is the Fracton-ator, which in fact is the simplest of all sampling schemes known in Stereology and for that very reason also the most powerful. Except by deliberately cheating, it is not known how to make a

biased fractionator! When estimating the number of particles in a pathological specimen, it is usually impractical to count them all. Instead the particle number is estimated in a known fraction of the reference space. The fractionator principle involves sampling particles uniformly at random with a known and predetermined probability, and then deriving the total number N in the reference space from the number in the sample and the sampling probability (Gundersen 1986, but the principle has been known outside the field of Stereology for a long time, see Jolly 1979). An organ containing particles is cut into pieces of unknown number, size and shape. A known and predetermined fraction of the pieces is sampled randomly (e.g. a fifteenth of all pieces is taken systematically with a random. start). Because any particle is contained in one of the segments, every particle has the same probability of being sampled, i.e. sampling is uniform . N can then be derived as the product of the particle number in the sample (N') and the inverse sampling probability: N= N' 15. This is an unbiased estimate of N. The sampling can be performed in two or more steps. In a large organ one may conveniently start by cutting it into a set of thick slabs, sample a fraction of them, cut these into bars, sample again from the bars, and cut the bars into small blocks of which a sample is then taken. This only takes a couple of minutes. The blocks are then embedded and sectioned exhaustively. However, one need sample only a small fraction of all sections (and a look-up section for each sampled section). In the last step one may then sample every tenth field of vision in every tenth row in a systematic design and in these fields count nuclei or glomeruli seen in one section and not in the other. The total number of counted particles times all the inverse sampling fractions is an unbiased estimate of N. If in the above steps a tenth to a hundredth was taken at each step one only counts from ~ 1/100,000 to ~ 1/ 1,000,000,000 of the particles - that is not a lot of work even in a whale kidney! The most remarkable feature of the fractionator is, that any need for dimensional information, like section thickness or magnification, is eliminated and the estimate is therefore independent of tissue deformation like shrinkage or swelling- even differential shrinkage does not influence the unbiased nature of the estimator. Note that in the discetor one estimates the sampled fraction = Vv ( dfis ) and shrinkage of the specimen re
( )

866

THE NEW STEREOLOGICAL TOOLS

mates in animal and human kidneys. Both series showed a CE of ~ 11 per cent, based on a mean glomerular count of 135 and 107 per kidney, respectively. The CE is therefore almost as low as one could expect it to be, under independence and uniformity the expected CE on a count of 121 is 1 ~ 121 = 9.l per cent. There are 600,000 and 23,000 glomeruli per kidney in man and rat, respectively (Nyenguard & Bendtsen 1989a and 1989b). THE SELECTOR ESTIMATING THE VOLUME OF PARTICLES AND ITS VARIATION IN A DISECTOR OF UNKNOWN THICKNESS The selector is the first of a series of techniques by which particle size and number may be estimated if just the magnification is known. It is based on a combination of the disector and the point-sampled intercepts and was first described by CruzOrive in 1987 in a paper which provided much impetus for the nucleator. Its principle is deceptively simple. In a stack of sections higher than the highest particle but otherwise of unknown and possibly varying section thickness, the first two sections are used as a disector for sampling n particles. These particles are followed through all the next sections, projected onto a systematic set of points, and complete intercepts are measured through every point hitting a sampled particle, see also the descriptions in Cruz-Orive 1987 and Gundersen 1986. All n sampled particles must be hit with a test-point at least once; if more than one intercept is measured in a particle one calculates the simple mean of the cubed lengths for that 3 particle, l 03, i . Since vi = (/3) l0 , i is an unbiased estimate of the volume of the i'th particle, it follows that
3 1 vN = n l0 ,i 3 i =1 n

Fig. 7. A rat kidney is cut into a series of slabs of roughly the same thickness using the "razor blade fractionator". The slab to be sampled first is found by means of random numbers. Here the aimed fraction is a third of the kidney. A random number between 1 and 3 is chosen (here no. 2). This slab and every third slab is sampled, here 2, 5, 8, 11, and 14. Statistically, this set of slabs contains on average exactly a third of everything in the kidney.

during dehydration and embedding after estimation of V(ref) may well be a problem for this estimate to be unbiased. See the discussion of this problem and a very simple elimination of it for small organs in Pakkenberg & Gundersen 1988. Whenever differential shrinkage might be a problem when comparing normal and diseased tissue or when paraffiin embedding is used, the fractionator may well be the simplest (and an excellent) solution to all such problems. Repeated estimates of the number of particles in the reference space will provide results varying around the true value. The efficiency of the fractionator (a low CE) depends on 1) homogeneity of the particle density in the reference space, 2) any random variation in the size of the pieces (the effect of their systematic variation is essentially limited by the systematic sampling), and 3) the number of pieces which are actually sampled (Gundersen 1986). The efficiency is improved by systematic sampling from a row of segments, see Fig. 7 and Gundersen & Jensen 1987a. The actual CE of a certain sampling scheme may be estimated rather easily, see Eq. 2.11 in Gundersen 1986 and the examples in Nyenguard & Bendtsen 1989a and l989b. In these studies the CE was estimated in groups of 8 and 25 systematically sampled, double esti-

is an unbiased estimate of the mean volume of the particles from the number distribution of particle volume: The particles were sampled uniformly in a disector. An example with rat Purkinje cells is shown in Table 2. Once the mean volume in the number distribution is known, a range of other estimates are posible. With an estimate of the reference volume V(ref), from either the Cavalieri-estimator or by just weighing the specimen, and of the volume
867

GUNDERSEN et al.

TABLE 2. Point-sampled intercepts in disector-sampled Purkinje cells followed on consecutive sections Cell no. Section no.
i 1 2 3 4 5 6 7 8 9 10 11 n = 11
3 l0 =
i=1 j=1 n i=1

2 30.7 30.7 30.7 68.0 41.1 41.1 30.7 7.9 30.7 2.5
3

k 1 3 4 5 3 1 4 3 5 4 4 37

j =1

l0 ,i

3 l0 ,i

68.0 21.8; 14.3 41.1 41.1;30.7 2.5;14.3 7.9 2.5 41.1 30.7 41.1 30.7 21.8 53.5 53.5 21.8 53.3

30.7

14.3

14.3 2.5 53.5

105.3 53.5 41.1

30.7 104.1 116.8 119.3 78.4 41.1 150.0 63.9 261.3 136.7 118.9 1221.2

30.7 34.7 29.2 23.9 26. 1 41.1 37.5 21 .3 52.3 34.2 29.7 360.7

l0,i k

= 1221.2 = 33.0u 3
37

n 3 3 l0,i = 1 l0,i = 1 360.7 = 32.8u 3 11 n i =1

CVN (v ) =

v v

V N

1 =

33.0 1 ~ 0.09 32.8

Eleven Purkinje cells sampled from an IUR stack of sections in rat cerebellar cortex. All cells seen in a random field of view in the second section but not in the first are sampled. On the second section and all following sections wherein the sampled cells (perikarya) are seen a set of systematic points is thrown at random. Through all points hitting the sampled cells the cubed length of the intercept l 03 , i is classified using the ruler shown in Fig. 6; these measurements are shown for each cell. A cell may be present in a section without being hit by a point and may also be hit by more than one point in a section. Each time a point hits a measurement is performed. The mean cubed intercept length 3 l 0 , i is calculated for each cell in the last column' the average of these 11 means times ~ is an unbiased estimator of 3 the mean Purkinje perikaryan volume: v N = 3 l 0 , i Taking the magnification and the ruler construction into account 3 the arbitrary ruler unit u shown here equals 71.5 m3, i.e. v N = 32.8 71.5 m3 ~ 2350 m3. The formula at the end of the table shows how to estimate the relative variation CVN(v) in the ordinary (number) distribution of perikaryan volume. Note that for this estimate one need not even know the magnification.

fraction VV(par/ref), from simple point-counting, one may estimate (indirectly) the total particle number by N(par) = V(ref) Vv(par/ref) / v N ( par ) still without knowing the section thickness. It is, however, a much more useful property of the selector that one may obtain a rather straightforward estimate of the relative variation in the number distribution of cell volumes, because both v N and v v are known, see Table 2. v v is estimated by the average of all measured intercepts, no matter which particles they come from, see the example in Table 2. It appears that Purkinje cells in adult rats
868

are relatively uniform in size with a CVN(v) ~ 0.1, but a sample of only 11 cells may be a trifle too small to be certain! The selector is a relatively efficient estimator of the variation in cell volume because the ratio between the two mean volumes v N / v v is estimated in the same cells. As in all other paired sampling designs, this eliminates a lot of sampling variation in the estimate. One should consider, however, that in most situations the disector is a more efficient estimator of particle numbers and the nucleator a more efficient estimator of size than the selector

THE NEW STEREOLOGICAL TOOLS

Fig. 8. Instrument for cutting small items macroscopically. The tissue is fixed to the horizontal, ruffled plate with 7 per cent agar. For optimal fixation the tissue must be put into the warm agar for a few minutes. The stability during cutting is best if the agar block is kept at ~ 5 C for some hours and cut while cold. The plate can be advanced between cuts at intervals of 1/6 mm using the threaded bolt at the end.

Fig. 9. Solid block of aluminium with semicircular groves of a depth of ~ 2 mm. When the block is kept at ~ 40 C for some time before the tissue is fixed to it with 5 per cent agar it is relatively easy to put the tissue rods parallel and with a random rotation for vertical sectioning.

for which a whole stack of serial sections has to be studied. The fact that one need not know the section thickness is of minor practical importance when using the optical disector. If cell volume variability is a parameter of major importance, it is probably more efficient to use Cavalieri's estimator on each cell in a stack of serial, optical sections (see the example with the nucleator in human cortical neurons below, and Fig. 4.1 in Gundersen 1986 for a very simple solution to the overprojection which occurs when transmission (projection) is used for the Cavalieri-estimator). THE NUCLEATOR ON VERTICAL SECTIONS When the nucleator is used on vertical sections it is possible to optimize the preservation of information about where the particles are sampled. This might, for example, be of importance when estimating the neuron perikaryon volume, since there is a pronounced variation in neuron size in the different layers of the cerebral cortex, and various disorders may well affect neurons in a particular layer. The example here is from the cerebral cortex of a rat, it is also used to illustrate some of the short-cuts possible in the method which reduce the workload considerably. For this reason a disector

with two physical sections was used. A rat brain hemisphere is embedded in agar and sliced into 1.5 mm coronal slices (Fig. 8). The volume of the cortex is estimated according to the Cavalieri-principle, see APMIS-1, using a test-system with a point-spacing of 1.0 mm at 20 X magnification using a dissection microscope (Olympus VE3 ). Each slice is cut in 1.5 mm wide, parallel rods, which are roughly perpendicular to the pial surface, approximately 50 rods in all. With a random start every sixth piece is sampled and embedded in agar in the indentations on the device shown in Fig. 9, with all pial ends in the same direction. Each rod is randomly rotated around its longitudinal axis, see also Fig. 8 in APMIS-1 . The agar block is then dehydrated and embedded in glycolmetachrylate and two pairs of 3.5 ,um sections from the middle of the block are Giemsastained. The section pairs are viewed in two microscopes projecting the images onto the table at ~ 1600 X. Neuronal nuclei which have their nucleolus in an unbiased frame in one section but are not present in the other section are sampled. Due to the vertical section design, directions in which to measure must be sine-weighted using the frame shown in Fig. 6. The vertical axis is clearly identifiable as the axis of the rods. The neuron nuclear and perikaryon volumes are estimated as
871

GUNDERSEN et al.

shown in Figs. 4 and 6. The mean volume of neuronal nuclei is ~ 500 m3 and that of neuronal perikarya is ~ 1040 m3. A surprisingly close relationship is present between the volume of individual neuron perikarya and the volume of their nuclei, r = 0.95 (Mller & Gundersen 1989). All the neuronal nucleoli have the same chance of being sampled with the disector, regardless of variations in size (height). There is, in fact, some variation in nucleolar size, the largest nucleoli are seen in the largest neurons which also have the largest nuclei. On an absolute scale the diameter variation is ~ 1 m which is not very large compared to the section thickness of ~ 3 m. The effect of selecting in one section all neurons which showed their nucleolus was therefore studied. As expected, the latter method leads to an overestimate of perykaryon volume: the largest neurons have a greater chance of being sampled. However, the bias was only ~ 3 per cent. This means that for a number of studies one might just sample neurons in one, independent section and estimate their volume in the same section with a very small bias, a bias which can be quantitated in a moderate number of animals. This is a great advantage when using stains which do not penetrate the thick plastic sections necessary for the optical disector. For a detailed description of this and a range of other time saving short-cuts, see Mller & Gundersen 1989 and Gundersen 1988. ESTIMATING THE SIZE DISTRIBUTION OF CELLS WITH THE NUCLEATOR IN AN OPTICAL DISECTOR Point-sampled intercepts provide unbiased estimates of the size (volume) of each sampled particle. The necessary uniform position of the sampling point does, however, mean that the estimate in a single particle is quite "noisy wherefore the 3 distribution of the observations, 3 l 0 , in practice is a useless estimate of the distribution of the particle volumes. Estimates made with the nucleator are different. The unique point in the particle, often the nucleolus in the nucleus, may be quite centrally located, and if the cell is not too oddly shaped, the nucleator-estimate of individual cell size may be a realistic estimate (and is unbiased in the mean). 3 4 The distribution of the observations, 3 l n , will then be a reasonable approximation of the real distribution of cell size. In order to establish this for a
870

particular cell type one must study the variation of the individual nucleator-estimates of size (volume or surface area) with respect to the true size of the same cells, a procedure which requires serial sectioning and the use of Cavalieri's estimator for individual cell volume and a spatial, linear test-system on IUR sections for surface, see Sandau 1987. When these two different approaches to size estimation are used on the same material it is possible to compare the efficiency of individual estimators, see Evans et al. 1989, and Evans & Gundersen l989b where neurons in the human cerebral neocortex have been studied this way. Due to the layered organization of neurons, sampling was carried out on sections perpendicular to the layering in order to reduce the sampling variance, resulting in a vertical section design for estimates of neuron (perikaryon) volume. However, should the vertical section direction deviate from perpendicularity to the layering, how will this effect the variance of the estimator? When neurons were sampled perpendicularly, at 60 , at 30 , and parallel to the layering, respectively, no differences were found in the variation of the nucleator estimate of neuronal volume. Another problem studied was the number of directions in which to measure the nucleator distance: two directions are always used, but would not 3, 4, or 6 directions be much better? Not surprisingly, there was very little difference in the variation of the nucleator estimate of volume when 3, 4 or 6 systematically selected, IUR directions were used. Even more directions are unlikely to be worth the effort as they will all lie in one plane (as long as only a central disector in an ordinary light microscope is used for observations). From a practical point of view, it is therefore recommended that 4 directions be measured for estimation of distributions. Note that on vertical sections both the two pairs of opposite directions must be chosen with the direction-finder shown in Fig. 6: the first (systematic) random, the next by adding 48 to the number of the first, otherwise the mean is not unbiased, see Evans & Gundersen l989b for further details. The coefficient of variation associated with a "4-way" nucleator is close to 1.0 when comparing for each neuron the estimate with the (much more elaborate) Cavalieri-estimate obtained from an exhaustive series of optical sections. This is quite acceptable when the coefficent of variation is compared to that of the size distributton of neurons in human temporal neocortex

THE NEW STEREOLOGICAL TOOLS

Fig. 10. The neuronal pericaryon volume distribution from the temporal cortex of an 80-year-old man, as estimated by the "4-way" nucleator. The mean neuronal pericaryon size is 1650 m3 (arrow) and a total of 48 cells were measured using a ruler similar to the one shown in Fig. 4, but the widths of the first and the last class were in the ratio 1:50. All layers were sampled uniformly, which is one reason for the relatively large variation in size.

Fig. 11. Diagram of a transverse section of the dentate gyrus in a rat brain. The cell bodies of the mononucleated granule cells are confined to one layer, the granule cell layer (g). The synapses on the dendrites of the granule cells are found in the molecular layer (m). The rectangle on the right shows the outline of a sample for EM.

which is about 1.2 to 1.4, see Fig. 10 and Gundersen 1988. This means that the observed variance of the size distribution is a few percent larger than the real variance: the added variance is 1 ~ n ~ 0.01 when 100 observations are made. It should be pointed out that the mean of the observed distribution is the correct mean and that the alternative for volume estimation is serial sectioning, which even with optical, serial sections represents a large amount of work. THE DOUBLE DISECTOR The following is an example of how one can avoid the non-trivial problem of estimating the thickness of ultrathin sections used in electron microscopy when using the disector to count objects that are small enough to require such thin sections. The specific example used for illustration involves estimating synaptic contacts in the dentate gyrus

of the rat brain. The idea is to estimate the numerical density NV of both synapses and neurons in the same series of ultrathin sections. The ratio of these densities, i.e. the number of synapses per neuron, may then be combined with an estimate of the absolute number of neurons at the light microscopic level to estimate the absolute number of synapses in the region under study. Earlier descriptions of this principle involving the ratio of small to large objects can be found in Gundersen 1986 and Brndguard & Gundersen 1986. In the dentate gyrus the neuronal cell bodies are confined to a single layer and their dendritic processes and synapses with afferent fibre systems confined to the adjacent molecular layer, see Fig. 11. The dentate gyrus is cut into 200 to 500 m thick parallel slabs in the horizontal plane of the brain. Alternate slabs provide material for the Cavalieri-estimate of the reference volume and for the estimate of the synapse to neuron ratio at the
871

GUNDERSEN et al.

Fig. 12. Diagram summarizing the double disector technique used to estimate the synapse to neuron ratio in a sample as shown in Fig. 11. Serial sections (k + I in number) are collected from a block containing a sample of the dentate gyrus. The first and last sections (a and b ~ are used for a disector estimate of the numerical density of neuron nuclei, N/neu/granJ, in the granule layer. Randomly sampled, adjacent sections within the series are used for the disector estimates of the synapse numerical density in the molecular layer, Nv~syn/molJ. Because of the large number of synapses on the dendrites of the granule cells in the molecular layer (m ), only a fraction of the area of the molecular layer is sampled in the three small unbiased sampling frames as shown.

electron microscope level. The blocks to be used for EM are trimmed from the slabs so that the deep and superficial borders and the arbitrarily defined lateral extents of the layer to be sampled can be observed when ultrathin sections removed from the block are viewed with low power EM. For the estimation of the synapse to neuron ratio in stratified tissue like the dentate gyrus one has to deal with the problem that the neurons and the synapses are in two different reference spaces, the granular and the molecular layers, respectively. It is therefore necessary to estimate the volumes of these two reference spaces separately, which causes no real difficulty. If the two densities are then estimated from uniform samples of the whole of the two reference spaces, granular and molecular layer respectively, the estimate of the number of synapses per neurons is unbiased:1
N(syn) N(neu) Q (syn) V ( mol ) a(syn) t Q (neu) V ( gran ) a(neu) k t Q (syn) a(neu) k Q (neu) a(syn) V ( mol )

V ( gran )

Ratio-estimators with a random variable in the denominator are generally not unbiased, but since the actual bias decreases much faster with increasing number of observations than the CE (Jensen & Gundersen 1982) they are considered UFAPP: unbiased for all practical purposes.

where t is the unknown thickness of the ultrathin sections, which cancels nicely in the equation, and k is the known ratio between the number of sections in disectors for neurons and synapses. In designs such as this example, k equals the number of sections in the disector minus one. a(syn) and a(neu) is the area of the frame used for counting synapses and neurons, respectively. The optimal distance between the sections used in the disectors is 1 to 2 m for granule cell nuclei. When a series of relatively thick (gold, ~ 100 nm) ultrathin sections is used, so that fewer sections are required to obtain the optimal height in the disector, this means that every 10th to 20th section can be used for the neuron disectors and that one or two random pairs of consecutive sections can be used for the synapse disectors, see Fig. 12. An estimate of the ratio of synapses to neurons made from a series of sections taken from one block is illustrated by the data in Table 3. A series of 15 sections was used. Corresponding fields of vision on adjacent sections of three systematic, randomly selected areas of the molecular layer were used for the synapse disector samples. The analyses were performed in both directions by interchanging the roles (sampling and look-up sections) of the top and bottom sections in the disectors, see Fig. 13. This doubles the sample size with very little extra

872

THE NEW STEREOLOGICAL TOOLS

TABLE 3. Hippocampal granule cells and synapses sampled in a double disector

Disector sample Granule cells Synpases, I Synapses, II Synapses, III Synapses, Total Q-(top) 22 9 8 12 Q-(bottom) 21 18 17 16

43 27 25 28 80

a(neu) = 0.013 mm2 a(syn) = 5.42 10-5 mm2

V(mol) 11.47 mm 3 = = 3.72 V(gran) 3.08 mm 3 N(syn) Q (syn) a(neu) k V(mol) = N(neu) Q (neu) a(syn) V(gran) 80 2 0.013 14 = 3.72 = 7700 43 6 0.0000542
Neurons and synapses counted in the granular layer of rat hippocampus using double disectors as shown in Fig. 12. Synapses are counted in three frames, each of area a(syn) = 5.42 10-5 mm2. The ratio of the two reference volumes are from Coleman et al. 1987.

effort and the same particle cannot thereby be counted twice! In practice, one should probably have at least two to four disectors uniformly sampled in the reference space, more if inhomogeneity is detectable in the sample. In situations where the study concerns localized changes due to local stimuli such a simple and strong design is unfortunately not possible. This raises the tricky problem of the correspondence: are the synapses in the molecular layer just above a certain region of the granular layer the synapses from the neurons in that particular region of the granular layer? Equally critical: is the delineation by two parallel lines perpendicular to the pial surface as in Fig. 11 a proper definition of the correspondence? There are no general answers to these questions, great caution must be exercised in order to avoid the "reference trap" of which some classical examples are given in Brndgaard & Gundersen 1986. See also the recent work by Cadete-Leite et al. 1988 who drew conclusions from estimates of the density of neurons in the granule layer in the hippocampus. Their conclusion (drawn in the title of the paper) was warranted only because they also measured the thickness of the granule cell layer, i.e. they had a reasonable measure of the marked change in the volume of the reference space.

Even after uniform sampling, the synapse to neuron ratio may not always be an adequate descriptor. In situations in which one cannot safely assume that the number of neurons is constant, conclusions drawn from differences in the synapse to neuron ratio may not be valid. When synaptic changes may occur and there is not adequate time for neuronal changes, such as the acute effecs of drugs or physiological stimuli, it may be valid to make assumptions about the absence of changes in neuronal number and the synapse to neuron ratio can be a valuable parameter. Similarly, in situations in which one wishes to investigate distributional differences, such as gradients in the innervation or localized effects of lesions or transplants, synapse to neuron ratios may clearly be of some value despite difficulties in general in interpreting the data rigorously. ANOTHER DOUBLE DISECTOR IN A PHYSICAL-OPTICAL COMBINATION What is described here is a principle which allows the unbiased estimation of all cells (endothelial, epithelial, mesangial and - if present - inflammatory cells) in an average glomerulus without regard to shrinkage and section thickness. The basic
873

GUNDERSEN et al.

874

THE NEW STEREOLOGICAL TOOLS

Fig. 13. Low power electron micrographs of the granule cell layer of the 1 st and 1 5th sections in the series with outlines of the nuclei of the -granule cells in the interposed drawings. The nuclei which appear in one section and not in the other are in black: Q = 22 in the upper section and 21 in the lower, see Table 3. The blood vessels are hatched. Bar is 10 m.

principle is to estimate separately the numerical density of cells in the glomeruli NN(cells/glom) and the mean glomerular volume v (glom). By definition
N N ( cells( glom ) = NV ( cells / glom ) v( glom )

The containing or reference space for Nv(cells/ glom) must be measured in the same units as v (glom) both formally and actually. The simplest and best way to do this is to make both estimates in the same sections. The units may then be

arbitrary and unknown, and shrinkage and other dimensional deformation of the glomeruli or of the cells become irrelevant. The estimation of NN(cells/glom) can be performed in kidney biopsies as well as in material from whole kidneys. Slightly different stereological methods, but based on the same principle, are used for plastic and paraffin embedded material, see Marcussen et al. 1989. For illustration, estimation of NN(cells/ glom) in autopsy kidney tissue embedded in paraffin is briefly described, the method also applies to biopsies with only trivial modifications.

Fig. 14. Two optical section planes through a glomerulus with a distance of approximately 3 m. The distance is illustrated by the thickness of the separation line between the two pictures. To the left is the top look-up plane and the nuclei clearly seen here are not counted. Moving 3 m towards the bottom five new nuclei come into focus in the central rectangle and are counted (right). A part of the randomly thrown tessellation of counting frames each of an area 18 m 28 m = 504 m2 is seen. Two unbiased counting frames in each glomerulus were used for counting through the whole section thickness. Hematoxylin-Giemsa stained. 875

Four slices of a few mm thickness were systematically sampled from a formalin fixed autopsy kidney. The slices were embedded in paraffiin and two Hematoxylin-Giemsa stained ~ 25 ,um thick sections were made from each. A glomerulus was defined as the minimal convex figure enclosing the glomerular tuft. NV(cells/glom) was estimated in an optical disector. The glomeruli were randomly sampled by moving perpendicularly from the capsule through the cortex to the medulla and back again using a stage drive. The fields of vision were projected onto the table using a projection microscope, see Fig. 3, and every fifth glomerular profile was chosen for subsampling at a high magnification using a 100X oil immersion objective with 1.40 NA. On the table a tesselation of counting frames, each of area a1, was placed randomly. Using a table of random numbers, two test frames were counted in each selected glomerulus. The look-up plane was defined as the two-dimensional section plane where the nuclei were first seen clearly when focusing from top to bottom in the section. All nuclei which were not intersected by the look-up plane were counted, see Fig. 14. A total of approximately 150 to 200 cells were counted in roughly 15 glomeruli. The average numerical density of cells in glomeruli is now NV(cells/glom) = Q-(cells)/(t a1) where t is the unknown thickness of the section. Using the pair of ~ 25 m thick sections as a physical disector one estimates
v(glom) = VV (glom / cortex) P (glom / cortex) = P NV (glom / cortex) NV (glom / cortex)

caps" were a minor problem when estimating glomerular volume (see Mller 1986 for another estimator of glomerular number using just fixed tissue cut in a vibratome). A stage drive was used to move the unbiased counting frame of area a2 through the cortex from the capsule to the medulla. 10X objectives and two fields of vision from two identical projection microscopes were used. As before NV(glom/cortex) = Q-(glom)/(t a2) where t is the same unknown thickness of the section which therefore cancels when combining the above equations
N N (cells / glom) = Q (cells) a2 PP (glom / cortex) Q (cells) a1

where VV(glom/cortex) is the volume fraction of glomeruli in the cortex estimated by point counting using a 10X objective on one section of the disector-pair and NV(glom/cortex) is the numerical density of glomeruli in the cortex. In a pilot study it was found that, at this magnification and in such thick sections overprojection and so-called "lost

The last term is, of course, just the ratio between the two reference volumes as in the double disector for estimating the number of synapses per neuron. A normal human glomerulus contains ~ 2500 cells. If it is also necessary to know the number of each type of cells in the glomeruli, differential counting has to be performed. This is best done in thin sections in a physical disector using e.g. silver staining as shown below for the follicle in Fig. 15. The main disadvantage of the method is that one has to make certain that "lost caps" are not a problem because the entire thickness of the section is used as the sampling space for the disectorcounting of cells in glomeruli. If nuclei have dropped out of the bottom plane of the section, one underestimates the numerical density, a phenomenon which does not affect counting in a disector which employs two section planes (Sterio 1984) or in an optical disector with measured height (Gundersen 1986) where one, for precisely that reason, avoids the bottom of the section. The magnitude of the bias is not known, but considering that ~ 25 m sections were used it is unlikely to be of much practical importance. In plastic sections it is probably not present at all.

Fig. 15. Two adjacent glycolmethacrylate sections of 1 m thickness through a mouse ovary. The orientation of the section pair is made isotropic through the use of Mattfeldts Orientator (See APMIS-1). The pair is selected from a number of parallel sections because the lower section is the first to hit the nucleolus of the oocyte. A quadratic test system is placed arbitrarily on the section with a corner on the nucleolus. In a systematic random quarter of all frames all granulosa cell nuclear profiles are sampled obeying the unbiased 2-dimensional counting rule, see Fig. 2. The 20 profiles through nuclei which are not seen in the upper section are encircled on the lower section. The staining used is a silver stain developed by Danscher 1983. The distance between the test lines is 34 m at the tissue level, the width of the white strip separating the micrographs is 1 m. Notice that the nuclei are not rarely separated by less then 1 m, a feature which makes it difficult to identify them unambiguously and therefore sample them unbiasedly using a disector of two physical sections. Using optical disectors this is, however, not a problem at all. F, O, and N indicate the follicles, limit towards the theca, the oocyte, and the oocyte nucleus, respectively. 876

THE NEW STEREOLOGICAL TOOLS

877

THE NUMBER OF BALLS IN A NUCLEATED BAG

This final example is the only one in this review where some critical information depends on knowledge of the section thickness. As illustrated by most of the above examples, a lot of effort and ingenuity (e.g. the Selector!) have been put into the development of methods over the last four years which eliminate the need to know the disector height, h. Another example of this is the procedure described in Pakkenberg & Gundersen 1988, where a disector-Cavalieri combination eliminates section thickness. As mentioned above, the use of optical disectors does, however, to a very large extent solve the problem in general. Fig. 15 shows an ovarian follicle from a mouse. To reproduction-biologists such a section represents a large amount of potential information: what are the volumes of the oocyte, its nucleus, the antrum, i.e. the excentric cavity (cavities?) and of the whole follicle? These answers can be obtained in a straightforward manner with the nucleator, since the follicle has an easily recognizable and unique unit in the middle, the nucleolus of the oocyte. Unbiased estimates of these volumes in this particular follicle are: v(follicle) = 5.16 106m3, v(oocyte) = 0.33 106 m3, v(oocyte nucleus) = 9930 m3, v(antrum) = 0.36 106m3. In order to obtain these estimates, the section which happens to hit the nucleolus was selected from a stack with isotropic orientation. This also means that most other stereological estimators cannot be used because they require the section to be uniformly random in position - and the particular section which hits the oocyte nucleolus is definitely not in a random position with respect to the well-ordered little universe of a follicle. Using the two consecutive sections shown in Fig. 15 as a disector of height h = 1 m we may, for example, think about estimating the number of granulosa cell nuclei in the sampled space of volume = h (follicle cross sectional area). Sampling in a systematic, random quarter of the frames, there are 20 nuclei hit by the section but not by the previous one, i.e.
NV (gran. cells / follicle central volume) ~ 20 7 ( 15000 )2 440 = 2.46 10 3 m 3

figure of ~ 2.5 granule cell per 1000 m3 is not the numerical density NV (gran.cells/follicle) of granulose cells in the follicle - because the disector is not uniformly positioned in the follicle. Then, what is the total number of granulosa cells in the follicle and how can we estimate that number by just observing one central and thereby arbitrarily positioned disector? The number of granulosa cells is clearly related to the total number counted in the disector, i.e. 4 Q- = 4 20, a number which is in turn directly proportional to the height h = 1 m of the disector: if h were e.g. 2 m, for example, we would expect Q- to be 160, and so on. The problem is, that since the disector has an isotropic orientation and passes through the nucleolus, a granulosa cell close to the nucleolus has a greater chance of being sampled (and counted) in the disector than another cell which is more distant from the nucleolus. The probability that the i'th granulosa cell is counted in the disector is therefore proportional to h and inversely proportional to the distance di from the nucleolus (the centre for the disector-rotation in space) to the granulosa cel1 nucleus:
Prob(counting granulosa cell) = 1 h d1 2

This is a unique feature: after having counted the i'th granulosa cell we can estimate its a priori probability of being captured by the probe (the disector) from measurements made in the probe itself! No other case of this is known in Stereology. It is now but a simple matter of inverting the above probability and summing over all n cells counted to provide an unbiased estimator of the total number of granulosa cell nuclei in the follicle under study:
N ( gran.cells / follicle ) = 2 d1
i =1 n

2 4 0.757 m 10 6 = 13,700 1m 440

where 7 is the number of upper frame corners hitting the follicle, the frame area is 15 by 15 mm, and the magnification is 440X. Unfortunately, this
878

The factor of 4 is related to the sampling scheme in which only a quarter of all cells in the disector pair were sampled in the example in Fig. 15. The constant 2 is specific for isotropically oriented sections. If vertical sections are used, the constant is , and di is the shortest distance from the i'th granulosa cell to the vertical axis (which passes through the nucleolus). For proofs of the above relations and a more detailed description of the method, see Bagger et al. 1988.

THE NEW STEREOLOGICAL TOOLS

Like the other estimators based on the nucleator principle the estimator of total number of particles per universe is an unbiased estimator independent of the shape of the universe, the position of the unique point, and of the shape, size and position of the particles to be counted. The complete cross section of the universe must of course be observable in the section where the unique point is seen. On the other hand, the statistical variation of the estimate, CE, clearly depends on all these factors and is also likely to vary inversely with the absolute number of granulosa cells sampled in the disector. Only the last factor is known and leads to an 1 estimated CE = 20 ~ 0.22 which is then a lower limit for the real CE in the example. If we want a more precise estimate we may sample all cells in the central disector instead of a quarter of them and, moreover, use a greater height in the disector. There is, however, likely to be a narrow limit as to how efficient these improvements really are: they do not decrease the contribution to the real CE from all the other factors mentioned above. Also, taking into consideration the biological variation between different follicles at the same stage of development and the fact that scientific statements therefore must be based on a sample of such follicles, it is obvious that an optimal sampling strategy involves a certain number of follicles and only a relatively small number of measurements in each. As opposed to more measurements per follicle, more follicles (n, say) in the sample reduces the contribution of all sources to the CE of 1 the estimated mean by a factor n . The optimal division of the sampling effort in follicles at varying stages is currently under study, see Bagger et al. 1988. Finally, according to Evans' idea (Evans & Gundersen 1989a) the section pair in Fig. 15 also contain all the necessary information to obtain an estimate of the 3-dimensional spatial distribution of all elements of the follicle with respect to the -oocyte nucleolus. There are, for example, a rather large number of granulosa cells in various stages of mitosis. Are-these cells uniformly distributed among the granulosa cells- in the 3-dimensional space of the follicle or is it, for example, mainly the most peripheral cells which are dividing? The unique feature of the nucleator which we can exploit here is that the 3-dimensional distance from the oocyte nucleolus to everything in the follicle is observable in the section. It is therefore

not only possible to estimate correctly the fraction of cells in mitosis but also the variation of the as a alunction of the mitotic index # cells distance,

Fig. 16. The variation in the local numerical density of glia cells as a function of the distance from a neuron nucleolus. Thirty neurons were sampled from human temporal cortex in isotropic sections and a total of 145 glia cells were counted in a design as shown in Fig. 15 but in optical disectors using the microscope in Fig. 3 (Evans & Gundersen 1989a). The graph clearly illustrates that there is a lower limit to how close glia cells can be to a neuron nucleolus - the radius of the neuron! - and that just outside the neuron there is a concentration of glia cells which is about five times higher than the overall numerical density of ~ 1.1 glia cell per 10.000 m3. Outside this crowd of glia cells around a neuron there may even be a zone at about a distance of 20 m with a particuIarly low density of glia cells.

i.e. its spatial distribution. Fig. 16 illustrates the analogous function of the numerical density: of glia cells in relation to neurons in human neocortex; this can be compared to Fig. 4. Pathologists like oceanographers and mining engineers never forget that their objects are 3-dimensignal - words like intercellular space, invasion, and hypervascularization lose their meaning if

applied to the section rather than to the cells and the tissue. The purpose of this review has been to show that when diagnosis or research needs to sharpen its aim by quantitating the maybe subtle 3-dimensional changes which might be decisive then there are handy and well understood instrumeets already developed for precisely that purpose.
The dedicated technical assistane of S. Bergman, M. D. Christiansen, D. Da Silva, E. Kabell-Kjr, A. Larsen, M. Lundorff; A. Nielsen, L. Nielsen, P. E. Nielsen & P. Schirmer is thankfully acknowledged.

Evans, S. M. & Gundersen, H. J. G.: Estimation of spatial distributions using the nucleator. Acta Stereol. 8: 395-400, 1989. Evans, S. M., Sandau, K., Jensen, E. B. & Gundersen, H. J. G.: A comaprison of two unbiased estimators of individual particle surface area. To be published. Gundersen, H. J. G.: Stereology of arbitrary particles. A review of unbiased number and size estimators and the presentation of some new ones, in memory of William R. Thompson. J. Microsc. 143: 3-45, 1986. Gundersen, H. J. G.: The nucleator, J. Microsc. 151: 3-21, 1988. Gundersen, H. J. G. & Jensen, E. B.: Particle sizes and their distributions estimated from line- and pointsampled intercepts. Including graphical unfolding. J. Microsc. 131: 291-310, 1983. Gundersen, H. J. G. & Jensen, L. B.: Stereological estimation of the volume-weighted mean volume of arbitrary particles observed on random section. J. Microsc. 138: 127-142, 1985. Gundersen, H. J. G. & Jensen, E. B.: The efficiency of systematic sampling in stereology and its prediction. J. Microsc. 147: 229-263, 1987a. Gundersen, H. J. G. & Jensen, E. B.: Estimation of moments of size distributions of arbitrary particles. Acta Stereol. 6: 197-199, 1987b. Gundersen, H. J. G., Bendtsen, T. F., Korbo, L., Marcussen, N., M11er, A., Nielsen, K., Nyengaard, J. R., Pakkenberg, B., Srensen, F. B., Vesterby, A. & West, M. J.: Some new, simple and efficient stereological methods and their use in pathological research and diagnosis. APMIS 96: 379-394, 1988. Helander, K., Hofer, P.-A. & Holmberg, G.: Karyometric investigations on urinary bladder carcinoma, correlated to histopathological grading. Virchows Archiv A. 403: 117-125, 1984. Howard, V.: Rapid nuclear volume estimation in malignant melanoma using point sampled intercepts in vertical sections: A review of the unbiased quantitative analysis of particles of arbitrary shape. In: Mary, J. Y. & Rigaut, J. P. (Eds.): Quantitative Image Analysis in Cancer Cytology and Histology, Elsevier Science Publisher, New York 1985, pp. 245-254. Howard, V., Reid, S., Baddeley, A. & Boyde, A.: Unbiased estimation of particle density in the tandem scanning reflected light microscope. J. Microsc. 138: 203-212, 1985. Jensen, E. B. & Gundersen, H. J. G.: Stereological ratio estimation based on counts from integral test systems. J. Microsc. 125: 51-66, 1982. Jensen, E. B., Kieu, K. & Gundersen, H. J. G.: On the stereological estimation of moment measures. Ann. Inst. Stat. Math. In press, 1990. Jolly, G. M.: Sampling of large objects. In: Cormack, R. M., Patil, G. P. & Robson, D. S. (Eds.): Sampling Biological Populations. Int. Co-operat. Publish. House. Maryland 1979, pp. 193-201. Marcussen, N.: Unbiased stereological estimation of the number of particles inside other particles. Submitted 1990. Michel, R. P. & Cruz-Orive, L.-M.: Application of the Cavalieri principle and vertical sections method to

REFERENCES
Bagger, P. V., Bang, L., Christiansen, M. D., Gundersen, H. J. G., Kabell-Kjr, E. & Mortensen, L.: Total number of particles in a bounded region estimated directly using the nucleator: Granulosa cell number in ovarian follicles. To be published, 1990. Bendtsen, T. F. & Nyengaard, J. R.: Unbiased estimation of particle number using sections - a historical perspective with special reference to the stereology of glomeruli. J. Microsc. 153: 93-102, 1989. Brngger, A. & Cruz-Orive, L. M.: Nuclear morphometry of nodular malignant melanomas and benign nevocytic nevi. Arch. Dermatol. Res. 279: 412-414, 1987. Brndgaard, H. & Gundersen, H. J. G.: The impact of recent stereological advances on quantitative studies of the nervous system. J. Neurosci. Meth. 18: 39-78, 1986. Brndgaard, H., Evans, S. M., Howard, C. V. & Gundersen, H. J. G.. Total number of neurons in human neocortex unbiasedly estimated using optical disector. J. Microsc. 157: 285-304, 1990. Cadete-Leite, A., Tavares, M. A. & Paula-Barbosa, M. M.: Alcohol withdrawal does not impede hippocampal granule cell progressive loss in chronic alcoholfed rats. Neurosci. Lett. 86: 45-50, 1988. Calverley, R. K. S. & Jones, D. G.: Determination of the numerical density of perforated synapses in rat neocortex. Cell. Tiss. Res. 248: 399-407, 1987. Coleman, P. D., Flood, D. G. & West, M. J.: Volumes of the Components of the Hippocampus in the Aging F344 Rat. J. Comp. Neurol. 266: 300-306, 1987. Cruz-Orive, L.-M.: Particle number can be estimated using a disector of unknown thickness: The seletor. J. Microsc. 145: 121-142, 1986. Dansher, G.: A silver method for counterstaining plastic embedded tissue. Stain Technology. 58: 356-372, 1983. Diem, K. & Lentner, C.: Documenta Geigy. Scientific Tables. J. R. Geigy, Basel 1970.

878

THE NEW STEREOLOGICAL TOOLS

lung: Estimation of volume and pleural surface area. J. Microsc. 150: 1 17-136, 1988. Mulvany, M. J., Baandrup, U. & Gundersen, H. J. G.: Evidence for Hyperplasia in Mesenteric Resistance Vessels of Spontaneously Hypertensive Rats Using a Three-Dimensional Disector. Circ. Res. 57: 794800, 1985. M11er, A., Strange, P. & Gundersen, H. J. G.: Efficient estimation of cell volume using the nucleator and the disector. J. Microsc. In press, 1990. M11er, J. C.: Dimensional changes of proximal tubules and cortical capillaries in chronic obstructive renal disease . Virchows Arch. A. 410: 153 -158, 1986. Nielsen, K., Colstrup. H., Nilsson, T. & Gundersen, H. J. G.: Stereological estimates of nuclear volume correlated with histopathological grading and prognosis of bladder tumour. Virchows. Arch. (Cell. Pathol.) 52: 41-54, 1986. Nielsen, K., Berild, G. H., Bruun, E., Jrgensen, P. E. & Weis, N.: Stereological estimation of mean nuclear volume in prostatic cancer, the reproducibility and the possible value of estimations on repeated biopsies in the course of disease. J. Microsc. 154: 63-69, 1989. Nielsen, K., rntoft, T. & Wolf; H.: Stereological estimates of nuclear volume in non-invasive bladder tumors (Ta) correlated with the recurrence pattern. Cancer. 64: 2269-2274, 1989. Nyengaard, J. R. & Bendtsen, T. F.: Glomerular number in a wide range of animals estimated by a simple and unbiased stereological method. To be published. Nyengaard, J. R. & Bendtsen, T. F.: Number and size of glomeruli, kidney weight and body surface in normal human beings. To be published.

Ooms, E. C. M., Blok, A. P. R. & Veldhuizen, R. W.: The reproducibility of a quantitative grading system of bladder tumours. Histopathology 9: 501-509, 1985. Pakkenberg, B. & Gundersen, H. J. G.: Total number of neurons and glia cells in human brain nuclei estimated by the disector and the fractinator. J. Microsc. 150: 1-20, 1988. Petran, M., Hadrarsky, M., Egger, M. D. & Galambos, R.: Tandem-Scanning Reflected-Light Microscope. J. Opt. Soc. Am. 58: 661-664, 1968. Sandau, K.: How to estimate the area of a surface using a spatial grid. Acta Stereologica. 6: 31-36, 1987. Sterio, D. C.: The unbiased estimation of number and sizes of arbitrary particles using the disector. J. Microsc. 134: 127-136, 1984. Srensen, F. B.: Stereological estimation of nuclear volume in benign melanocytic lesions and cuteneous malignant melanomas. Am. J. Dermatopathology. 11: 517-527, 1989a. Srensen, F. B.: Objective histopathological grading of cutaneous malignant melanomas by stereological estimation of nuclear volume: Prediction of survival and disease-free period. Cancer 63: 1784-1798, 1989b. Srensen, F. B.: Stereological estimation of mean nuclear volume from vertical sections with analysis of size variability. J. Microsc. In press, 1990. Vesterby, A.: Unbiased estimation of the number of osteoclasts in iliac crest biopsies. To be published. The list of references is updated in this 2nd printing in 1990, but the references in the text are unchanged.