Professional Documents
Culture Documents
Abstract 1. Introduction
In this work we have proposed an automatic Imaging plays an important role in diagnosing of
method to analyze the MRI head scans and detect brain tumors. Magnetic Resonance Imaging (MRI) is
abnormality in brain due to tumors. We make use of well suited for monitoring and evaluating cerebral
the bilateral symmetry of the human brain. We use tumors. In MRI, they appear either as hypo-intense
T2 axial scans as they are highly sensitive to (darker than brain tissues) or as iso-intense (same
pathological process. Tumors appear as hyper intensity as brain tissue) in T1-weighted scans and
intense in T2 scans and have intensity close to that as hyper-intense (brighter than brain tissues) in T2-
of cerebrospinal fluid (CSF). For normal slices, CSF is weighted scans [1, 2]. The perifocal edema also
symmetrical about the vertical central line. So the appears as hyper-intense in T2-weighted scans.
presence of abnormal tissues in the CSF class can be Since T2–weighted images are very sensitive in
detected by measuring the vertical symmetry of the detecting brain pathology, patients with suspected
CSF image. Our method has four parts. First, the intracranial disease should be first screened with
brain is extracted from head using brain extraction T2-weighted spin-echo and FLAIR images [3]. If an
algorithm (BEA) and used to detect the boundary abnormality is found then additional scans are
between the cerebral hemispheres (CH). Next a slice taken to characterize the lesion. As per the brain
transformation has been done to align it with the screening protocol suggested by Hasselink [3], T1
world space for bilateral symmetry checking. Then, a images need to be taken only if T2 images show
fuzzy segmentation is done to generate a CSF abnormalities. Tumors appear as hyper-intense in
image. Finally a fuzzy symmetric measure (FSM) is T2 scans and have intensity close to that of CSF. In
calculated for CSF image to discriminate between normal slices, the CSF is symmetric about the
normal and abnormal scans. Experiments using our vertical central line. Therefore, if any abnormal
method were done on 20 volumes of normal and tissue is present in the CSF, the symmetry is
abnormal datasets. Two measures, false alarm (FA) disturbed. This can be detected by measuring the
and missed alarm (MA) were used to quantify the symmetry of the CSF about the vertical central line
performance of our method and found to be very [1].
less, 3% and 6% respectively. These measures were Several methods have been developed to automate
also used to compare our method with that of the brain tumor localization, analysis, segmentation
existing methods. The mean FA is lower than the and visualization [1, 4-18]. Khotanlou et al., [4] and
existing methods. However the MA is higher, and is Wang et al, [5] have given a summary of existing
due to the tumors present symmetrically about the methods and discussed their strengths and
vertical central line. weaknesses. Some of the methods are semi-
automatic and require user interaction either to
Keywords: hyper intense, bilateral symmetry, brain initiate or adjust the result [9-11]. Some methods
extraction, fuzzy segmentation, fuzzy symmetric are atlas-based [14-16] and require predefined atlas.
measure, false alarm, missed alarm. Few methods require multi-channel images (T1, T2,
1
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
PD and contrast enhanced images) [1, 17, 18] which checking is done about the boundary line
are not always available in clinical routine. Some separating the CHs. Liang et al [19] defined that the
methods use symmetry property of brain tissues to boundary between the CHs must lie within the
segment the tumor [4, 5, 12]. Some methods use interhemispheral fissure (IHF). IHF is a deep cleft
features like edges, gray values and local contrast of and a longitudinal fissure between the CH and is
pixels within a block for identifying tumorous filled by CSF. Our method detects the boundary
characteristics [6-8, 13]. The block-based analysis is that corresponds to IHF using the contour of brain.
a time consuming process for a decision support We used our brain extraction algorithm (BEA) [20]
system. to extract the brain from T2 MR scans and is named
Generally, most tumor segmentation methods have as T2-BEA. T2-BEA perfectly extracted the entire
two stages. Stage one is initialization that roughly brain even if they are separated by fissures [20]. It
locate the tumor and the second stage refine the also eliminated the non-brain tissues like eye and
rough result and produce the whole tumor. scalp even if they have the similar intensities to
Khotanlou et al., [4] proposed a 3D method for T1 brain tissues. The performances of the T2-BEA is
images. They use two alternate approaches to found to be better than the well known existing
detect the initial position of the tumor in stage-1. BEAs, brain extraction tool (BET) [21] and brain
The first one is a fuzzy classification method that surface extractor (BSE) [22]. Using the brain contour,
classifies the brain into 5 classes, CSF, WM, GM, an intelligent system is developed to join the edges
tumor and background. Then a sequence of of the cleft that corresponds to IHF present in the
morphological operations is performed to classify contour [23]. Our system worked successfully on
the tumor from the tumor class. The second several normal and abnormal 2D slices. This system
approach is based on symmetry analysis of gray will be useful for further processing like tissue
level histograms of normal hemisphere and segmentation, image compression, registration [24]
pathological hemisphere. In their method a user and bilaterally symmetry analysis procedures [23].
interaction is required to select the pathological The detected boundary is then aligned with central
hemisphere. The fully automatic fuzzy classification vertical line for symmetry checking. The symmetry
method is applicable only to hyper-intense nature of the CSF segmented from the brain portion is
of tumors whereas the semiautomatic symmetry checked by using fuzzy symmetry measure (FSM).
analysis method is applicable to tumors that are The remaining part of this paper is organized as
localized in one hemisphere only. In stage-2, a follows. In section 2, the materials used in our
deformable model to refine the initial tumor experiment are given. In section 3 an overview of
segmentation is used. A tumor detection method the method used is discussed. The results and
developed by Wang et al [5] is also based on discussions presented in section 4. The conclusion is
symmetry property of the brain. It uses a symmetry given in section 5.
measure, calculated between the gray level
histogram of the two cerebral hemispheres (CH), to 2. Materials
classify the slice as normal or abnormal. Both the Twenty datasets of normal and abnormal subjects
aforesaid methods are applicable to cerebrum were used in our experiments. The details of the
tumors and use the gray level histograms of the CHs datasets are given in Table 1. The first ten datasets
for symmetry analysis. Both of the methods need (v01-v10) consisting of 2 normal volumes and 8
either a user interaction or employ a complex abnormal volumes with brain tumor (Neoplastic
function to initialize the process. disease) and multiple sclerosis were taken from the
In this paper, we present a knowledge-based website ‘The Whole Brain Atlas’. The remaining ten
paradigm that combines the feature extraction, datasets (v11-v20) were taken from KGS Scan
image processing algorithms, graphics and fuzzy Centre, Madurai, India, which were acquired from
techniques to produce an unsupervised system 1.5T Siemens machine for 8 normal subjects and 2
capable of detecting the abnormality automatically abnormal subjects affected by tumors. The
in T2-weighted MRI scans. If an asymmetry is found dimension of each slice in volumes v01-v10 is
in a slice then there is more likely a tumor to be 256 × 256 pixels and slice thickness varying from 2-
present in it [12]. In axial orientation, the upper 5mm with 260mm field of view. So the pixel
slices corresponding to cerebrum are approximately dimension is fitted to 1 × 1 mm. The slices taken
symmetrical bilaterally. Hence our algorithm is from 1.5T Siemens machines have a dimension of
suitable to the cerebrum abnormalities, especially 448 × 512 pixels and 5mm thickness (except v18:
caused by single tumor. The bilateral symmetry 3mm) with 40-50% inter-slice gap. The field of view
2
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
is fixed to 210 × 240 mm and hence the pixel and GM regions. The presence of abnormal tissues
dimension is set to 0.47x0.47 mm. Naturally, the in the CSF class can be detected by measuring the
voxel dimension is in anisotropic form in all our vertical symmetry of the CSF image. FSM is used as
datasets. a threshold to discriminate between normal and
abnormal CSF images. The Y axis of the Cartesian
Table 1: Details of Datasets used coordinate system is used as the vertical axis with
Data Volume Total
Gender Age Clinical origin at the centre of each slice.
set Identity slices
1 v01 Female 81 Normal 54
Our method comprise of a sequence of four
2 v02 Female 76 Normal 43 processes as given below.
3 v03 Female 51 Anaplastic 56
Astrocytoma 1. Brain extraction
4 v04 Male 35 Astrocytoma 29
2. Slice transformation
5 v05 Male 62 Metastatic 24
adeno 3. Tissue segmentation
carcinoma 4. Abnormality checking
6 v06 Female 42 Metastatic 24
bronchogenic 3.1. Brain Extraction
carcinoma
7 v07 Male 75 Meningioma 27
Numerous brain extraction algorithms (BEA) are
8 v08 Male 22 Sarcoma 24 available in literature [25, 26]. Our T2-BEA [20]
9 v09 Male 30 Multiple 24 makes use of anisotropic diffusion process [27],
Scelerosis optimal thresholding and morphological processes
10 v10 --- --- Multiple 54
[28] to separate the brain from non-brain portions.
Scelerosis
11 v11 Female 18 Normal 19
The diffusion process is used to highlight the brain
12 v12 Female 19 Normal 19 from T2 head scan. Then an intensity threshold is
13 v13 Male 43 Normal 19 computed using which a rough binary brain portion
14 v14 Female 32 Normal 19 is generated. The morphological operations, erosion
15 v15 Male 45 Normal 19
and dilation, and connected component analysis are
16 v16 Male 51 Normal 19
17 v17 Female 39 Normal 19
then performed on the rough brain portion to
18 v18 Male 43 Normal 19 produce the brain mask. Finally the brain mask is
19 v19 Female 55 Tumor 19 used to extract the brain from T2 scans.
20 v20 Male 38 Tumor 19 Then we find the boundary between CHs that
corresponds to IHF. An intelligent system is used
3. Method Overview [23] to detect the edges of the IHF present in the
In MR images, some of the non-brain tissues like fat, extracted brain. In each 2D axial slice, edge1 is
skin, muscles and background clutters have nearly marked at the nadir of the upper portion of cleft
the same intensity as that of tumors. Therefore the and edge2 is fixed at the zenith of the lower portion
brain portion should be extracted first to eliminate of cleft. Then a line is drawn by joining these two
these non-brain tissues. The extracted brain contour edges that corresponds to the boundary line
is used to detect the line separating the cerebral passing through IHF.
hemispheres. This linear boundary corresponds to
IHF. When an MRI is acquired, the head position 3.2. Slice Transformation
sometimes may not coincide with the world co- The detected boundary line is used to transform the
ordinate of the image acquisition system. There may MRI scans to the ideal world co-ordinate, here the
be a tilt in the acquired image. To perform the Cartesian co-ordinates system. The line
symmetry test for the CSF in MRI, the tilt has to be corresponding to IHF is aligned with the vertical
corrected. For this a slice transformation is done. central line, Y-axis of the Cartesian coordinate
The structural arrangement of the whole brain is system. This alignment is done by composite
similar on the both side of hemisphere and is transformation of translation followed by rotation.
symmetrical about the vertical axis through the The mapping procedure (M) to move from the
brain center. Tumors appear hyper intense in T2 imaging coordinates (IC) to world coordinates (WC)
scans and have intensity close to that of CSF. The is given by:
transformed brain is segmented into four major
classes WM, GM, CSF and background. This MWC, IC = R(90-θ) . T(d)
segmentation is useful to retain the cluster (1)
containing CSF and tumor and to remove the WM
3
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
90-θ
θ
Figure 1: (a) MR scan with detected boundary (in dotted white color) and Cartesian coordinate system with centre as origin (in light gray
color) (b) Transformed image with coordinates θ and d
cluster j, xi is the ith element of d-dimensional updating the cluster centers and the membership
measure data, cj is the d-dimensional center of the grades for each data point, FCM iteratively moves
cluster and dij is the Euclidean distance between ith the cluster centers to the “right” location within a
data point (xi) and jth centroid (cj). data set.
The fuzzy portioning is carried out through an
iterative optimization of the objective function given The FCM algorithm as proposed by Bezdek [32] is:
in equation (3), with the update of membership uij
and the cluster centers cj using:
4
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
Step 1: Randomly initialize the membership calculated and compared with the existing method
matrix U = [uij], U(0) that has a constraint predictive cognitive system for brain images (PCB).
equation given by Three successive versions of PCB, v.1, v.2 and v.3 [6-
c 8] were used to compare our results.
∑u
i =1
ij = 1, ∀j = 1,..n , (6)
5
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
5. Conclusion
We have developed a fully automatic system which
can rapidly classify the given MRI data sets into
normal and abnormal. Our method is purely based
Figure 3: CSF image with CSF regions (white) and central vertical on intensity and is applicable to MR images with
line (light grey) for symmetry checking single mass effect present within a hemisphere. Out
of 8 abnormal data sets our method was able to
The qualitative validation of our method was done detect the abnormality in 6 and produced false
by four experts. The outputs produced by T2-BEA, alarm only for 2 data sets, when compared to 4 in
standard space alignment, segmentation of CSF PCB. This work is currently extended to segment the
regions and their symmetry comparison along tumors and to compute the volume.
central vertical line were approved by the experts.
They further suggested that identifying the Acknowledgement
abnormality in MRI head scans might be useful to The authors wish to thank Dr. K.G. Srinivasan M.D.,
quicken the diagnosis and prognostic processes for R.D., Consultant Radiologist, KGS Advanced MR &
single tumor or mass effect present in either of the CT Scan, Madurai, Tamilnadu, India and Dr. N.
hemispheres. Karunakaran DMRD., DNB., Consultant –
The quantitative parameters, FA and MA, calculated Radiodiagnosis, Meenakshi Mission Hospital and
for 2 normal and 6 abnormal datasets are given in Research Centre, Madurai, Tamilnadu, India for
Table 2. FA for the proposed system is obtained providing the MRI Head scans and for giving the
only for v04 and v06 due to the swelling edema. qualitative validation. The authors would also wish
This effect narrowed down the sulci at higher levels to thank Dr. K. Selvamuthukumaran M.Ch. (Neuro),
and thus reduced the cerebral CSF as shown in Sr. Consultant, Department of Neuro Surgery,
Fig.4(a) and (b). This is also reflected in FSM of CSF Meenakshi Mission Hospital and Research Centre,
and recognized them as abnormal slices. Missed Madurai, Tamilnadu, India and
alarm is obtained for v05, v06, v07 and v08. In v05, Dr.S.P.Balachandran,M.D.,D.M., (Neuro), Neurologist,
the lesion is characterized by low signal on T2 image Dindigul Neuro Centre, Dindigul District, Tamilnadu,
with very little surrounding edema (Fig.4(c)). Hence India for their help in verifying the results.
they are missed by the FSM calculation. In v06, the
end of lesion is represented by small region and This work is catalysed and funded by Science for
thus missed by FSM. The lesion in v07 and v08 is Equity, Empowerment and Development (SEED)
equally scattered on both sides as shown in Fig.4(d) Division, Department of Science and Technology
and thus missed by the FSM calculation. None of (DST), Government of India, New Delhi, Grant
the alarm has been noticed for the normal volumes. number: SP/YO/011/2007 under the Scheme for
The values of FA and MA calculated by our method Young Scientists and Professionals (SYSP).
along with that of the PCB are given in Table 3. The
FA indication in our method is lower than all the
6
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
Figure 4: Slices corresponds to false alarm (a and b) and missed alarm (c and d) (a) slice from vo4, affected by intracranial hypertension
at right side of hemisphere (b) slice from v06, affected by intracranial hypertension at lower right side of hemisphere (c) slice from v05
with low signal tumor (d) slice from v07 with equally scattered lesion
Table 3: Comparison of FA and MA obtained in our method with that of PCB (v.1, v.2 and v.3)
Total FA (%) MA (%)
No. Volume
Slice PCB v.1 PCB v.2 PCB v.3 Proposed PCB v.1 PCB v.2 PCB v.3 Proposed
1. v01 54 11 11 0 0 - - - -
2. v02 43 10 23 0 0 - - - -
3. v03 56 0 5 0 0 24 0 0 0
4. v04 29 25 8 7 13 48 5 5 0
5. v05 24 37 27 17 0 50 17 11 13
6. v06 24 5 5 0 13 28 0 0 4
7. v07 27 19 15 7 0 33 3 0 7
8. v08 24 14 19 9 0 25 4 4 25
Average 15 14 5 3 26 4 3 6
References [5] Wang Z., Hu Q., Loe K., Aziz A., and Nowinski
[1] Clark M.C., Hall L.O., Goldgof D.B., Velthuizen R., W.L., Rapid and Automatic Detection of Brain
Murtagh F.R., and Silbiger M.S.,Automatic Tumors in MR Images, Proceedings of SPIE,
Tumor Segmentation using Knowledge-Based Bellingham, WA, vol., 5369, pp. 602-612, 2004.
Technique, IEEE Transaction on Medical Imaging, [6] Lau P.Y., and Ozawa S., PCB: A Predictive System
vol. 17, no. 2, pp. 187-201, 1998. for Classifying Multimodel Brain Tumor Images
[2] Thamburaj A.V., Neurology and Systemic in an Image-Guided Medical Diagnosis Model,
Malignance, Apollo Hospital, Chennai,India, Proceeding of the 12th International Conference
http://www.thamburaj.com/cns_systemic_malign on Intelligent System for Molecular Biology,
ancy.htm, 2009. Glasgow, UK, 2004.
[3] Hesselink J.R., Basic Principles of MR Imaging, [7] Lau P.Y., and Ozawa S., A Region- and Image-
Department of Radiology, University of Based Predictive Classification System for Brain
California, Tumor Detection, Proceeding of Symposium on
http://spinwarp.ucsd.edu/NeuroWeb/Text/br- Biomedical Engineering, Hokkaido, Japan, pp.
100.htm, 2009. 72-102, 2004.
[4] Khotanlou H., Colliot O., Atif J., and Bloch I., 3D [8] Lau P.Y., and Ozawa S., A Multiparameter
Brain Tumor Segmentation in MRI using Fuzzy Hierarchical Representation using Region-Based
Classification, Symmetry Analysis and Spatially Estimation Model For Detecting Tumor in T2-
Constrained Deformable Models, Fuzzy Sets and Weighted MRI Brain Images, Malaysian Journal
Systems, vol. 160, pp. 1457-1473, 2009.
7
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010
Of Computer Science, vol. 18, no. 1, pp. 1-19, [20]Somasundaram K., and Kalaiselvi T., Fully
2005. Automatic Brain Extraction Algorithm for Axial
[9] Jiang C., Zhsng X., Huang W., and Meinel C., T2-Weighted Magnetic Resonance Images,
Segmentation and Quantification of Brain Submitted after II revision to Computers in
Tumor, IEEE International Conference on Virtual Biology and Medicine, 2010.
Environment, Human-Computer Interfaces and [21] Smith S.M., Fast Robust Automated Brain
Measurement Systems – VECIMS 2004, Boston, Extraction, Human Brain Mapping, vol.17, pp.
USA, PP. 12-14, July 2004. 143-155, 2002.
[10] Kaus M.R., Warfield S.K., Nabavi N., Black P.M., [22] Shattuck D.W., Sandor-Leahy S.R., Schaper K.A.,
Jolesz F.A., and Kikinis R., Automated Rottenberg D.A., Leahy R.M., Magnetic
Segmentation of MR Images of Brain Tumors, Resonance Image Tissue Classification Using a
Radiology, vol. 218, pp. 586-591, 2001. Patrial Volume Model, NeuroImage, vol. 13(5),
[11] Xie K., Yang J., Zhang Z.G., and Zhu Y.M., Semi pp. 856-876, 2001.
Automated Brain Tumor and Edema [23]Somasundaram K., and Kalaiselvi T., A Novel
Segmentation using MRI, European Journal of Technique for Finding the Boundary between
Radiology, vol. 56, pp. 2219-2233, 2003. the Cerebral Hemispheres from MR Axial Head
[12] Mancas M., Gosselin B., and Macq B., Fast and Scans, Proceeding of The 4th Indian
Automatic Tumoral Area Localization Using International Conference on Artificial
Symmetry, Proceedings of IEEE ICASSP Intelligence – IICAI-09, Tumkur, Bangalore, India,
Conference, Philadelphia, Pensylvania, USA, pp. 597-603, December 16-18, 2009.
2005. [24] Zaidi H., Montandon M.L. and Slosman D.O.,
[13] Dubey R.B., Ratan R., Hanmandlu M., and Gupta Magnetic Resonance Imaging-Guided
S.K., Computer Assisted Segmentation of Brain Attenuation and Scatter Corrections in Three-
Tumor, TechnoramA, A Supplement to IEI News, Dimensional Brain Positron Emission
pp. 23-26, March 27, 2008. Tomography, Medical Physics, vol. 30, no. 5, pp.
[14] Cuadra M.B., Pollo C., Bardera A., Cuisenaire O., 937-948, 2003.
Villemure J., and Thiran P., Atlas Based [25] Segonne F., Dale A. M., Busa E., Glessner M.,
Segmentation of Pathological MR Brain Images Salat D., Hahn H. K. and Fischl B., A Hybrid
using a Model of Lesion Growth, IEEE Trans in Approach to the Skull Stripping Problem in MRI,
Medical Imaging, vol. 23, no. 10, pp. 1301-1313, NeuroImage, vol. 22, pp. 1060-1075, 2004.
2004. [26] Somasundaram K., and Kalaiselvi T., Brain
[15] Moon N., Bullitt E., Leemput K.V., and Gerig G., Extraction Method for T1 Magnetic Resonance
Model Based Brain and Tumor Segmentation, Scans, Accepted in IEEE International
ICPR Quebec, pp. 528-531, August 2002. Conference on Signal Processing and
[16] Prastawa M., Bullitt E., Ho S., and Gerig G., A Communication - SPCOM2010, Indian Institute
Brain Tumor Segmentation Framework based on of Sciences (IISc), Bangalore, India, July 18-21,
Outlier Detection, Medical Image Analysis, vol. 2010.
18, no. 3, pp. 217-231, 2004. [27] Perona P., and Malik J., Scale-Space and Edge
[17] Capelle A.S. Colot O., and Fernandez C., Detection using Anisotropic Diffusion, IEEE
Evidential Segmentation Scheme of Multi-Echo Trans. on Pattern Analysis and Machine
MR Images for the Detection of Brain Tumors Intelligence, vol.12, no.7, pp.629-639, 1990.
using Neighborhood Information, Information [28]Sonka M., Hlavac V., and Boyle R., Image
Fusion, vol. 5, pp. 203-216, 2004. processing, Analysis and Machine Vision, II
[18] Dou W., Ruan S., Chen Y., Bloyet D., Constans Edition, Brooks / Cole Publishing Company,
J.M., A Framework of Fuzzy Information for the Pacific Grove, CA, 1999.
Segmentation of Brain Tumor Tissues on MR [29] Harris G., Andreasen N.C., Cizadlo T., Bailey J.M.,
Images, Image and Vision Computing, vol. 25, Bockhold H.J., Magnotta V.A. and Arndt S.,
pp. 164-171, 2007. Improving Tissue Classification in MRI: A Three-
[19] Liang L., Rehm K., Woods R., and Rottenberg D., Dimensional Multispectral Discriminant Analysis
Automatic Segmentation of Left and Right Method with Automated Training Class
Cerebral Himispheres from MRI Brain Volumes Selection, Journal of Computer Assisted
using Graph Cut Algorithms, NeuroImage, vol. Tomography, vol. 23, no. 1, pp. 144-154, 1999.
34, pp. 1160-1170, 2007. [30] Yang M.S., Hu Y.J., Lin K.C.R. and Lin C.L.L.,
Segmentation Techniques for Tissue
8
ICGST-GVIP Journal, Volume 10, Issue 3, August 2010