Professional Documents
Culture Documents
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 12, DECEMBER 2011
I. INTRODUCTION
Fig. 1. (a) Example retinal image in the STARE dataset (i.e., im0043.bmp)
[18]. (b) Corresponding intensity image.
LU: ACCURATE AND EFFICIENT OPTIC DISC DETECTION AND SEGMENTATION BY A CIRCULAR TRANSFORMATION
2127
2128
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 12, DECEMBER 2011
(2)
and
denote the L1-norm image grawhere
dients in the horizontal and vertical directions, respectively.
and denote the numbers of image rows and columns. The OD
probability map is determined as follows:
(3)
where the OD center can be detected at the brightest pixel
within
. For the example retinal image in Fig. 1(a)
and Fig. 3(a) shows the determined OD probability map.
In the proposed technique, the OD is detected by searching
from the first 20% brightest pixels within the OD probability
map shown in Fig. 3(b). Though the OD may not lie at the
brightest pixel within the OD probability map (92.6% over
STARE dataset as reported in [16]), it nearly always lies within
the first 20% brightest pixels within the OD probability map.
It should be noted that the search space can also be reduced
by the OD detected by other methods such as Youssifs [17].
In such case the reduced search space becomes a circle that
is centered at the detected OD center [17] with a radius much
larger than the OD radius, says, 90 pixels. Mahfouzs method
is used because it is ultrafast (0.6 s for the STARE dataset) and
the 20% brightest pixels within the OD probability map are
reliable to include the OD center properly.
B. Circular Transformation
A transformation is designed to capture the OD-specific
image and shape characteristics. The transformation is termed
circular transformation because it performs best while a
perfect circular region is present. The circular transformation
is designed based on the observation that for a point within a
(roughly) circular-shaped image region, the variation of the
distances from the point to the region boundary reaches the
minimum when the point lies exactly at the region centroid.
Specifically, for each retinal image pixel it first detects multiple
pixels with the maximum variation (denoted by PMs in the
ensuing discussion) along multiple evenly-oriented radial line
(4)
where
and
denote the positions
of two image pixels along the th radial line segment neigh. An image variation matrix
boring to
can thus be determined for the pixel at
. It should
where
be noted that the pixels are indexed from
is closer to
than
.
Therefore, if the pixel at
lies within the OD, the
evaluated image variation across the OD boundary is usually
positive as the OD is usually brighter than the surrounding.
, the positions of the
For the retinal image pixel at
PMs along the evenly-oriented radial line segments can be
denoted by an index vector as follows:
(5)
where
indicates the position of the PM along the th line
segment, i.e., the index of the maximum of the th row of
. The maximum image variation along the radial
line segments can therefore be denoted as follows:
(6)
LU: ACCURATE AND EFFICIENT OPTIC DISC DETECTION AND SEGMENTATION BY A CIRCULAR TRANSFORMATION
2129
where
gives the maximum image variation along
the th radial line segment. The distances from all PMs to the
under study can thus be determined by a dispixel at
tance vector as follows:
(7)
where
is the distance from the pixel at
to the
PM along the th line segment at
as follows:
(8)
For image pixels within the OD, many detected PMs lie exactly at the OD boundary due to the clear image variation across
the OD boundary. But some detected PMs may lie at other positions due to the retinal vessels/lesions. The PM outliers can be
filtered out based on a pair of OD-specific constraints. First, the
PMs with a zero/negative maximum variation are filtered out because for pixels within the OD, the maximum image variation at
the detected PMs is usually positive. The PMs with a zero maximum variation are often detected for pixels around the border
of the central circular retinal region where the radial line segments quickly probe into the surrounding dark region. The PMs
with a negative maximum variation are often detected for pixels
at some dark region such as the macula center where the image
consistently becomes brighter with its distance from the macula
center.
The remaining PMs are further filtered based on their dis. A distance threshold is defined as follows:
tances to
Fig. 5. OD map construction and PMs detection. (a) OD map of the example
retinal image in Fig. 1(a). (b) Close-up view of the PMs detected for the pixel
at the located OD center (labeled by the cross).
Fig. 6. OD boundary pixel detection. (a) Close-up view of the remaining PMs
after the two-stage filtering. (b) Close-up view of the OD boundary pixels created based on symmetry of the OD boundary with respect to the OD center: the
black points denote the remaining PMs shown in (a) and the white ones denote
the created symmetry-based OD boundary pixels.
(9)
where
is a median function.
denotes a
subset of
in (7) where some distance elements
(corresponding to the pixels with a zero or negative maximum
variation) have been removed in the first-stage filtering. As (9)
evaluates the deviation of
from
shows,
that usually approximates the OD radius
is at the OD center.
when
The circular transformation finally converts a retinal image
into an OD map as follows:
(10)
and
refer to the maximum variation
where
and distance of the PMs respectively [as defined in (6) and (7)]
denotes the mean of
after the two steps of filtering.
. Therefore, the numerator in (10) gives the overall
image variation and the denominator gives the standard deviation of
. The OD map can usually be enhanced by
incorporating the OD probability map as follows:
(11)
where
denotes the OD probability image in (3)
[shown in Fig. 3(a)]. For the example retinal image in Fig. 1
and Fig. 5(a) shows the final OD map computed.
2130
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 12, DECEMBER 2011
An OD boundary can thus be determined through B-spline fitting of the final OD boundary pixels. It should be noted that the
detected OD center and OD boundary pixels need to be mapped
back to the retinal images of the original resolution for proper
evaluation. For the example retinal image in Fig. 1(a) and (b)
shows the determined OD center and OD boundary of the original resolution.
D. Parameter Value Selection
The designed circular transformation has several parameters.
The first is the number of the radial line segments as described
in Section III-B. In general, has little effect on the OD detection when it is larger than 60. But it may affect the OD segmentation because the number of the PMs after the filtering may
not be enough to fit an OD boundary smoothly when is too
small. Experiments show that 180 radial line segments are sufficient for OD boundary fitting and the use of more does not
help much. Therefore, 180 line segments are used in the implemented system.
The second and also the most important parameter is the
length of the radial line segments. Generally speaking,
should be larger than the OD radius so that for retinal image
pixels around the OD center, the PMs along the radial line segments can be detected at the OD boundary. As the relative size
between the OD and the corresponding retinal image usually
varies within a specific range, parameter can be set based
on the diameter of the central circular region of the retinal
image as illustrated in Fig. 2(a). In the implemented system,
is set at
that is much larger than the OD radius that usually
and
(around
for most retinal
varies between
images). In addition, the use of a large also reserves the space
that excludes the ending section of each radial line segment (to
be discussed next).
Lastly, the performance of the circular transformation can
often be improved by excluding the starting and ending sections of each radial line segment from the image variation evaluation in (4). By excluding the starting section PMs will not
be detected at the retinal vessels (lying around the OD center)
for retinal image pixels lying around the OD center. By excluding the ending section PMs will not be detected around
the border of the central circular retinal region of the retinal
image where large image variation is often present due to different imaging artifacts such as uneven illumination. In the implemented system, the lengths of the starting and ending secthat is much smaller than the OD
tions are both set at
ending section does not
radius. It should be noted that the
of each radial line segment.
always correspond to the last
of each radial line segment
Instead it refers to the last
that lies within the central circular retinal region as illustrated
in Fig. 2(a) (that is how the border of the central circular retinal
region is excluded).
IV. EXPERIMENTAL RESULTS
This section presents experimental results. The three public
datasets used are first described. The performance of the designed circular transformation is then presented and discussed.
online
LU: ACCURATE AND EFFICIENT OPTIC DISC DETECTION AND SEGMENTATION BY A CIRCULAR TRANSFORMATION
2131
Fig. 7. OD detection and segmentation under the presence of retinal lesions: The three rows shows example retinal images, the converted OD maps, and the
close-up view of the finally fitted OD boundary, respectively. The bold white boundary in the third row denotes the manually traced reference OD boundary, the
thin black boundary denotes the OD boundary determined by the proposed technique.
Fig. 8. OD detection and segmentation under the presence of imaging artifacts: The three rows show example retinal images, the converted OD maps, and the
close-up view of the finally fitted OD boundary, respectively. The bold white boundary in the third row denotes the manually traced reference OD boundary, the
thin black boundary denotes the OD boundary determined by the proposed technique.
white curves and light black curves in the fifth row correspond
to the reference OD boundary and the one determined by the
circular transformation, respectively. As Figs. 7 and 8 show, the
proposed technique is tolerant to retinal lesions and imaging
artifacts and capable of locating both the OD center and the OD
boundary accurately.
Quantitative results show that an average OD detection
accuracy of 99.5% is obtained for 1401 retinal images within
the three datasets (i.e., 1197/1200 for the MESSIDOR dataset,
117/120 for the ARIA dataset, and 80/81 for the STARE
dataset) based on the first criterion. Most failure cases can be
explained by retinal lesions and imaging artifacts that almost
completely impair the OD-specific image and structural charac-
2132
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 30, NO. 12, DECEMBER 2011
TABLE I
COMPARISON OF THE OD DETECTION METHODS ON STARE DATASET (THE
SPEEDS OF FORACCHIA et al. [19] AND YOUSSIF et al. [17] ARE BOTH
TAKEN FROM MAHFOUZ et al. [16])
dius that is around 60 pixels for retinal images within the two
datasets. The small OD center error can be explained by the circular transformation that is designed to detect the OD centroid
exactly. At the same time, the six-pixel OD center error is much
smaller than that of state-of-the-art methods [13], [16][19] as
shown in Table I where the OD center error ranges 1429 pixels.
The large OD center error of the state-of-the-art methods can be
explained by the fact that the retinal vessels seldom converge to
the exact OD center. It should be noted that the OD center error
for the proposed method is evaluated on retinal images of the
original resolution.
Furthermore, the proposed technique is capable of segmenting the OD from many images of pathological retinas as
illustrated in Figs. 7 and 8. Based on the region overlapping
criterion, OD segmentation accuracies of 93.4% and 91.7% are
obtained for the 80 and 117 correctly detected retinal images
of the STARE dataset and the ARIA dataset, respectively. In
particular, imaging artifacts often have larger effects on the OD
segmentation compared with retinal lesions. This is because
retinal images with imaging artifacts often have low image
variation across the OD boundary where many PMs are not
detected as illustrated in the first three retinal images in Fig. 8.
On the other hand, retinal images with retinal lesions often
have relatively high image variation across the OD boundary
where PMs can be detected properly as illustrated in Fig. 7.
This also explains the higher OD segmentation accuracy of
STARE dataset that is composed of a large number of images
of pathological retinas with retinal lesions.
The direct comparison with state-of-the-art methods is difficult because most reported OD segmentation methods [4],
[22], [26][28] are evaluated on a private dataset. In the experiments, the active contour model technique [32] is implemented
and tested on the STARE dataset. Experiments show that the
OD segmentation accuracy is just around 70%. On the other
hand, the proposed technique achieves high OD segmentation
accuracy due to its flexibility. In particular, the circular transformation is capable of detecting certain proportion of OD
boundary pixels even though certain OD boundary segments
are severely degraded. But Hough transform and active contour
model both rely heavily on a closed shape model whose behavior varies greatly when certain OD boundary segments are
severely degraded. Second, the circular transformation is more
flexible to incorporate the OD-specific anatomical knowledge.
For example, it is very straightforward for the circular transformation to create the symmetry-based OD boundary pixels as
described in Section III-C.
C. Discussion
The proposed technique is efficient as shown in Table I. For
the STARE dataset, it takes around 5 s (implemented in Matlab
on a Dell PC with a 3.00 GHz CPU and 3.25 GB of RAM) on
average for both OD detection and OD segmentation. But for
most reported methods [13], [17][19], it takes 24.5 min to
perform the OD detection alone on similar computing platforms.
Mahfouzs method [16] is ultrafast but it only detects the OD
center with a low accuracy at 92.6% and an OD center error at
14 pixels. The high efficiency of the proposed technique is due
to the circular transformation as well as the two preprocessing
steps described in Section III-D, i.e., the image down-sampling
and the search space reduction.
In addition, the parameters of the circular transformation are
fixed as described in Section III-D for all experimental results
described in the last subsection. On the other hand, the proposed
technique is stable when the parameters vary within a certain
range. Fig. 9(a) and (b) show the performance variation with the
change of two key parameters, i.e., the number and the length
of the radial line segments. As Fig. 9(a) shows, the OD detection accuracy of the three datasets varies little when 30, 60, 90,
120, 180, 270, and 360 radial line segments are implemented
). The OD segmentation accuracy of
(where is fixed at
the last two datasets instead improves slightly when more radial
line segments are implemented but the improvement saturates
when is bigger than 180. In addition, the OD detection accuracy reduces when becomes too big or too small deviating
from the relative OD radius as shown in Fig. 9(b) (where is
fixed at 180). But the OD segmentation accuracy for all correctly detected retinal images varies little with the change of .
It should be noted that the lengths of the skipped starting and
because the OD radius
ending sections are both fixed at
.
is much larger than
Most OD segmentation error comes from three sources at
the current stage. First, it may be introduced for some special
retinal images that have much larger image variation across the
cup boundary than that across the OD boundary. Under such
circumstance, many PMs will be located at the cup instead of
the OD boundary improperly. Second, it may be introduced for
some special retinal images that have an ultra-low image variation across the OD boundary. For such retinal images, many OD
boundary sections have no PM detected and so the OD boundary
LU: ACCURATE AND EFFICIENT OPTIC DISC DETECTION AND SEGMENTATION BY A CIRCULAR TRANSFORMATION
2133