You are on page 1of 7

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES, VOLUME 4, ISSUE 1, JAN-JUNE /2017

ISSN (ONLINE): 2394-8442

SEGMENTATION AND ANALYSIS OF SCHIZOPHRENIC


MRI IMAGES
SHEREENA JOSE K[1], SUBHASHREE S[2], N. SRI MADHAVA RAJA[3]
[1, 2, 3]
Electronics and Instrumentation Engineering, St. Josephs College of Engineering, Chennai, India
[1]
shereenajose95@gmail.com, [2]subha9696@gmail.com, [3]nsrimadhavaraja@stjosephs.ac.in

ABSTRACT.

Enlargement of lateral ventricles is one of the consistent findings in schizophrenia. Segmentation of


medical images can be region-based or edge-based. This paper concentrates on obtaining regional information
like area and perimeter of lateral ventricles from segmented images and also carries out a comparison of few
edge detectors. Edge detection plays an important in medical image analysis. It can help to extract necessary
information from a raw image. Four edge detection methods namely Canny, Sobel, Prewitt and Roberts were
compared based on metrics assessing image quality. These techniques are applied to both healthy and
schizophrenic images and the best methods are determined based on Structural Similarity Index and Feature
Similarity Index. We aimed at finding detectors that could effectively trace the edges of lateral ventricles
specifically and the whole brain region in general.

KeywordsSchizophrenia; Lateral Ventricles; MRI; Edge Detection

I. INTRODUCTION
Schizophrenia is a severe mental disorder characterized by major symptoms that can be categorized into positive symptoms (e.g.
hallucinations and delusions), negative symptoms (e.g. social withdrawal and bizarre behavior) and cognitive symptoms (e.g. poor attention and
impaired thinking).It is a complex disorder with both genetic and environmental factors contributing [1]. No cure has been found yet, but it is
treatable through early diagnosis [2].

With the advent of CT and MRI, brain abnormalities have been confirmed in schizophrenia [3]. MRI studies in Schizophrenia concern
first-episode and chronic disease; though brain abnormalities are significant only in the chronic kind [4].Longitudinal changes in volume are
evident in hippocampus, cortical gray matter and temporal gray matter but not in thalamus [4]. One of the consistent findings is ventricular
enlargement. Enlarged ventricles inversely correlate with reduced mid-sagittal corpus callosum volume in first episode schizophrenia [5].
According to Gaser, the shrinking of the thalamus appears to be an important contributor to ventricular enlargement [6]. Increase of ventricular
volume is proven to be large in left temporal horn and is significant in males than in females [7].

Feature extraction techniques are important for analyzing images diagnostically. Edge detection serves the purpose of retaining important
structural properties of images while filtering out useless information. Edges are created in an image based on abrupt changes in intensity. Weak
edges can be cleaned up by increasing the threshold value. The quality of the edge is very much dependent on the image quality. Unlike Sobel
method, Canny method gives good edge detection by forming smooth edges [8]. A drawback of Canny is that it uses more memory space during
processing. Laplacian of Gaussian uses second order derivative to find edges and hence is sensitive to noise and cannot detect edge direction.
Morphological edge operations are apt for noisy MRI images when compared to techniques using first and second order derivatives [9].

II. METHODOLOGY
A. Image Acquisition

The imaging data were obtained from the Mind Clinical Imaging Consortium (MCIC) database [10] through the Mind Research Network
(www.mrn.org) via the Schiz Connect website (http://schizconnect.org). As such, the investigators within Schiz Connect [11] shared data but did
not participate in analysis or writing of this paper. The downloaded dataset comprised multiple imaging protocols from which only MPRAGE
images with T1 weighted dominance acquired with Siemens TRIO 3T scanner were used for analysis. The dataset contained images of both
Schizophrenia Broad patients and people of no known disorder. The DICOM images were converted to JPEG for ease of use in Mat lab using
the Irfan View software without compromising the quality. In this paper, we compare images of a schizophrenic patient and a healthy control
both aged 45.

To Cite This Article: SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA


RAJA,.SEGMENTATION AND ANALYSIS OF SCHIZOPHRENIC MRI IMAGES. Journal for Advanced
Research in Applied Sciences ;Pages: 312-318
313. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

B. Brain Region Extraction

Skull stripping [12] is mostly done by specialized algorithms that delineate the brain boundary. To begin with the grey-matter, white-matter,
and CSF masks should be kept and everything else must be removed. Then thresholding is done to separate the skull from the brain. Labeling is
done and the skull region can be isolated from the other brain regions or else everything except the skull can be retained. This leaves us to focus
our work only on the brain matter.

C. Regional Properties of the Lateral Ventricles

The skull-stripped images serve as the input to this stage. To extract the CSF regions from the whole brain matter, the following sequence of
operations are performed. First thresholding [13] is done by fixing optimal values for lower and upper thresholds based on the intensities of the
CSF regions. The output images obtained have many minute fragmented CSF regions. These are removed by applying morphological operations
[14], [15] like image dilation. The resulting images are then complemented to make the pixels in the CSF regions white and the remaining black.
This is done when algorithms consider pixels with value 1 as foreground pixels and pixels with value 0 as background pixels. Then the regional
descriptors like area and perimeter of the lateral ventricles are computed using the region props function in MATLAB.

Thresholding the Applying Computing Area


skull-stripped Morphological and Perimeter
Image Operations

Fig. 1. Sequence of Operations to get Area and Perimeter

D. Edge Detection

Segmentation can be based on similarity or discontinuity in pixel values. When similar pixels are grouped based on predefined criteria,
regions are created. When partitioning is based on discontinuities in intensity values, the segmentation is edge-based. Four edge detectors which
find edges by using first order derivatives are considered.

1) Canny

The steps in Canny edge detection are smoothing by Gaussian filter, finding the local gradient and edge direction, non-maximal suppression
and performing edge linking [15]. The local gradient is given by

The angle at which maximum rate of change occurs is

Equation (2) indicates edge direction. An edge point is a point whose strength is locally maximum in the gradient direction. Non-maximal
suppression is done by tracking tops of ridges that were created by edge points and by assigning zero to pixels that do not lie on ridge top so as
to thin the output. Ridge pixels are threshold using two values T1 and T2, where T2 is usually thrice of T1 [16]. Finally, weak edges are linked
to the strong ones if they show 8-connectivity.

2) Sobel

The Sobel detector uses vertical and horizontal masks to compute derivatives (gradient) separately for each direction. The gradient is given by

The magnitude of (3) is

The gradient magnitude describes how much neighboring pixels differ in intensity. The horizontal and vertical masks are given by

The algorithm involves filtering of images with both masks separately and adding them to finally compute the square root. Threshold can be
specified or can be found automatically. Edge pixels are those that have intensities greater than or equal to the threshold.
314. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

3) Prewitt

The Prewitt detector also uses masks to approximate the first derivatives and digitally. It has similar parameter values and is slightly
simpler to compute than the Sobel detector. The masks are given by

A drawback of Prewitt is that it tends to give somewhat noisier results [15].The coefficient 2 in Sobel offers smoothing.

4) Roberts

The Roberts detector is one of the oldest and simplest detectors. The masks are given by

Due to its asymmetry its functionality is limited.

E. Image Quality Assessment

Qualities metrics help evaluate the quality of processed images. Two images are taken for comparison-a reference image which is distortion-
free and a second image whose quality is to be measured. Both their matrix dimensions should match. Two popular metrics are used here.

1) Feature Similarity Index

FSIM algorithm [17] calculates the similarity index by considering the luminance component for gray scale images. The image should have a
dynamic range 0-255. It takes into advantage the complementary nature of phase congruency and gradient map. FSIM is based on the fact that
the Human Visual System concerns only with the low-level features of an image.

2) Structural Similarity Index

SSIM [18] index measures the extent to which two images are similar by comparing local patterns of pixel intensities normalized for
luminance and contrast. It brings image quality assessment from pixel based stage to structure based stage. Its value lies between -1 and +1.
SSIM is 1 for two identical images.

III. RESULTS AND DISCUSSION


Before the application of certain processing techniques, it is a normal practice to skull-strip the image to concentrate only on the brain
matter. So the skull is removed and only the brain matter is extracted from the original image. The T1-weighted coronal slices are shown in Fig.
2. Their corresponding skull-stripped images are shown in Fig. 3.
315. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

A. Regional Properties of the Lateral Ventricles

Thresholding of images is done by choosing an optimal lower and upper threshold based on the intensities in the CSF regions. Fig. 4 shows the
CSF portions and clearly the lateral ventricular CSF. The minute fragments are removed by using morphological operations and the resulting
output images are displayed in Fig. 5. Fig. 6 shows the complemented images as most algorithms consider pixels with value 1 as foreground
pixels and pixels with value 0 as background pixels. The approximate values of area and perimeter of the lateral ventricles are found by using
regional properties feature in MATLAB and the values are tabulated in Table I.

Fig. 4. Threshold Images

TABLE I. REGIONAL PROPERTIES OF THE LATERAL VENTRICLES

Image Area Perimeter


Abnormal 1667.0 291.7
Normal 976.0 264.4

B. Edge Detection

After skull-stripping, the edges are detected using the four methods discussed. The values of thresholds are adjusted such that the output is
optimal. Then these values are fixed for future edge detection purposes and no changes are made. The results are obtained by implementing all
the four techniques in a single program for both schizophrenic and healthy images. To inspect the efficiency of outputs for lateral ventricles, the
region is selected semi-automatically. Now the results are compared for the whole brain matter by using two similarity indices: SSIM and FSIM
and specifically analyzed for lateral ventricles by visual inspection. The visual comparison for lateral ventricles is first discussed. The Canny
method gives a smooth edge map but certain edge points in ambiguous places are not detected. Still the method detects weak edges better and is
least sensitive to noise. On examining the Sobel method the edges are quite continuous but a little rugged. This algorithm has given a good result
for smooth curves of the lateral ventricles in case of schizophrenic images. When it comes to thin and sharp edges in healthy images, the
algorithm is not that efficient in tracing out the edges. The results of Prewitt algorithm are nearly similar to those of Sobels but it is not suited
for noisy images. The Roberts algorithm for edge detection is very much sensitive to noise since the cross kernels used are small. But the edges
are fully covered in a rugged fashion.
316. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

The similarity indices are computed for schizophrenic and healthy edge detected output images separately as shown in Table II.

TABLE II. OUTPUT VALUES OF SIMILARITY INDICES

Images Metric Canny Sobel Prewitt Roberts


SSIM 0.6303 0.6633 0.6531 0.6486
Abnormal
FSIM 0.7406 0.7961 0.7875 0.7794
SSIM 0.7274 0.7444 0.7387 0.7373
Normal
FSIM 0.8636 0.8888 0.8889 0.8833
317. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

The values are computed comparing with the reference images generated by ImageJ software shown in Fig. 11

Fig. 11. Reference images generated by ImageJ

ACKNOWLEDGMENT
The imaging data was collected and shared by [University of Iowa, University of Minnesota, University of New Mexico, Massachusetts
General Hospital] the Mind Research Network supported by the Department of Energy under Award Number DE-FG02-08ER64581. Funding
was done by NIMH cooperative agreement 1U01 MH097435.

V. CONCLUSION
The lateral ventricle regions were analyzed for their regional properties. The approximated values of area and perimeter show that there is a
significant increase in the size of lateral ventricles in schizophrenic patients. The method is simple and assures ease of diagnosis. Four edge
detection techniques namely Canny, Sobel, Prewitt and Roberts were applied to both Schizophrenic and healthy images. Structural and Feature
Similarity indices were determined for edge detected images of both healthy and schizophrenic patients against reference images. Sobel and
Prewitt algorithms are found to show superior performance over the other two algorithms. This would help in future work involving analysis of
lateral ventricles in Schizophrenia and brain MRI in general.

REFERENCES
[1] Naoko Aoyama Glutamatergic Metabolites and Gray Matter Losses in Schizophrenia: A Longitudinal Study Using In Vivo Proton
Magnetic Resonance Spectroscopy, (2011) Electronic Thesis and Dissertation Repository

[2] Basic Facts About SCHIZOPHRENIA 14th edition (2008) , British Columbia Schizophrenia Society

[3] Martha E. Shenton et al. A review of MRI findings in Schizophrenia, Elsevier, Schizophrenia Research 49 (2001)

[4] Grant Steen R., Courtney Mull, Robert McClure, Robert M. Hamer and Jeffrey A. Lieberman Brain volume in first-episode schizophrenia
Systematic review and meta-analysis of magnetic resonance imaging studies, British Journal of Psychiatry (2006) , 1 8 8 , 51 0-51 8

[5] Elisabetta C. del Re et al Enlarged lateral ventricles inversely correlate with reduced corpus callosum central volume in first episode
schizophrenia: association with functional measures Brain Imaging and Behavior DOI 10.1007/s11682-015-9493-2

[6] Christian Gaser, Igor Nenadic, Bradley R. Buchsbaum, Erin A. Hazlett, Monte S. Buchsbaum Ventricular Enlargement in Schizophrenia
Related to Volume Reduction of the Thalamus, Striatum, and Superior Temporal Cortex (Am J Psychiatry 2004; 161:154156)

[7] T. Yotsutsuji et al Quantification of lateral ventricular subdivisions in schizophrenia by high-resolution three-dimensional magnetic
resonance imaging Psychiatry Research: Neuroimaging 122 (2003) 112

[8] Zolqernine Othman, Habibollah Haron, Mohammed Rafiq, Abdul Kadir Comparison of Canny and Sobel Edge Detection in MRI Images

[9] A.V.Kavitha, Madhulika.Y , Siddardha Kumar .M Comparison of Edge Detection Techniques for Brain MRI Image, IRACST -
International Journal of Computer Science and Information Technology & Security (IJCSITS), ISSN: 2249-9555 Vol. 5, No3, June 2015

[10] Gollub, RL, et al The MCIC collection: a shared repository of multi-modal, multi-site brain image data from a clinical investigation of
schizophrenia Neuroinformatics 2013 July ; 11(3): 367388. doi:10.1007/s12021-013-9184-3.

[11] Wang, Lei, Alpert, Kathryn I., Calhoun, Vince D., Cobia, Derin J., Keator, David B., King, Margaret D., Kogan, Alexandr, Landis, Drew,
Tallis, Marcelo, Turner, Matthew D., Potkin, Steven G., Turner, Jessica A., Ambite, Jose Luis, SchizConnect: Mediating Neuroimaging
Databases on Schizophrenia and Related Disorders for Large-Scale Integration, NeuroImage , doi:10.1016/j.neuroimage.2015.06.065

[12] https://in.mathworks.com/matlabcentral/fileexchange/25157-image-segmentation-tutorial

[13] https://in.mathworks.com/matlabcentral/fileexchange/29372-thresholding-an-image

[14] Hui-Fuang Ng, Cheng-Hung Chuang, Chih-Hsueh Hsu Extraction and analysis of structural features of lateral ventricle in brain medical
images, 2012 Sixth International Conference on Genetic and Evolutionary Computing
318. SHEREENA JOSE K, SUBHASHREE S AND N. SRI MADHAVA RAJA,.SEGMENTATION AND ANALYSIS OF
SCHIZOPHRENIC MRI IMAGES. Journal for Advanced Research in Applied Sciences; Pages: 312-318

[15] Rafael C. Gonzalez, Richard E. Woods, Steven L. Eddins Digital Image Processing using MATLAB 2004 Pearson Education, Inc.

[16] Yang Chao A comparison of medical image analysis algorithms for edge detection June 2010,Bachelor Thesis,Computer Science

[17] Lin ZhangaLei, ZhangaMember, Xuanqin Moub, and David Zhanga FSIM: A Feature Similarity Index for Image Quality Assessment,
IEEE Transactions on Image Processing (Volume 20, Issue: 8, Aug. 2011)

[18] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. SimoncelliImage quality assessment: From error visibility to structural similarity,IEEE
Transactios on Image Processing, vol. 13, no. 4, pp. 600-612,Apr. 2004. http://www.ece.uwaterloo.ca/~z70wang/research/ssim

You might also like