Computer Assisted Diagnostic System in Tumor Radiography

J Med Syst (2013) 37:9938 DOI 10.1007/s10916-013-9938-3

ORIGINAL PAPER

Computer Assisted Diagnostic System in Tumor Radiography Ahmed Faisal & Sharmin Parveen & Shahriar Badsha & Hasan Sarwar & Ahmed Wasif Reza

Received: 30 January 2013 / Accepted: 6 March 2013 # Springer Science+Business Media New York 2013

Abstract An improved and efficient method is presented in this paper to achieve a better trade-off between noise removal and edge preservation, thereby detecting the tumor region of MRI brain images automatically. Compass operator has been used in the fourth order Partial Differential Equation (PDE) based denoising technique to preserve the anatomically significant information at the edges. A new morphological technique is also introduced for stripping skull region from the brain images, which consequently leading to the process of detecting tumor accurately. Finally, automatic seeded region growing segmentation based on an improved single seed point selection algorithm is applied to detect the tumor. The method is tested on publicly available MRI brain images and it gives an average PSNR (Peak Signal to Noise Ratio) of 36.49. The obtained results also A. Faisal : S. Parveen Department of Computer System & Technology, Faculty of Computer Science & Information Technology, University of Malaya, 50603, Kuala Lumpur, Malaysia A. Faisal e-mail: [email protected] S. Parveen e-mail: [email protected] S. Badsha : A. W. Reza (*) Faculty of Engineering, Department of Electrical Engineering, University of Malaya, 50603, Kuala Lumpur, Malaysia e-mail: [email protected] A. W. Reza e-mail: [email protected] S. Badsha e-mail: [email protected] H. Sarwar Department of Computer Science and Engineering, United International University, Dhaka, Bangladesh e-mail: [email protected]

show detection accuracy of 99.46 %, which is a significant improvement than that of the existing results. Keywords Image denoising . Region growing . Segmentation . Tumor detection . MRI brain . Medical image processing

Introduction Image denoising and segmentation are the two essential but challenging tasks in image processing, especially while dealing with the medical applications, like detection of tumor growth from MRI (Magnetic Resonance Imaging) brain images. MRI is a very powerful tool to define the tumor region in the brain but suffers from noise and low contrast which in turns immensely affects the accuracy of segmentation. Over the last two decades, the rapid development in medical imaging technology with its various high precision imaging modalities, in parallel, the immense advancement in computer technology has enormously transformed the medical diagnostic process of healthcare sector. Medical imaging allows detection of tumor, cancer, fibroid, cyst, etc., in the human body. X-ray, Computed Tomography (CT) scans, Ultrasound, Positron Emission Tomography (PET), and Electrocardiogram (ECG), are widely and routinely used clinical practice. MRI is believed to be a potential diagnostic tool and medical imaging technique to visualize the detailed internal structures in radiology, hence giving an accurate organ anatomy [1]. The capacity of MRI to contrast between the different soft tissues has ensured wide use for the diagnosis of brain tissue abnormalities. It provides useful information about shape, size, and localization of tumors. Moreover, unlike other modalities, patient does not require to be exposed to a high ionization radiation. However, MRI images are corrupted by noise during image acquisition. Thermal noise and sample

9938, Page 2 of 10

resolution are the two examples of MRI noise sources. All these incorporated noises degrade the quality of the image and pose obstacles for effective computer-aided analysis through feature extraction, recognition, interpretation, and quantitative measurements. Medical image processing deals with facilitating accurate observations of abnormalities to the physicians for proper diagnosis of the disease. Identification of tumor from brain MRI images includes several steps of image processing. Among them, denoising and segmentation plays the vital role to detect the tumor from an enhanced image [2]. However, there is a contradiction between image denoising and enhancement. Denoising is the process of removing noise from a signal. Conversely, doing enhancement, followed by removing noise will make the noise amplified (making the denoising effect poor). Moreover, most existing segmentation algorithms tend to be very sensitive to noise [3]. For these reasons, it is necessary to first preprocess the MRI data by denoising. There are several algorithms proposed in the literature for denoising MRI data, namely, median filter, winner filter, Gaussian filter, and anisotropic-diffusion filter [4]. Experimental results show that the fourth order Partial Differential Equation (PDE) based technique is a good choice for denoising MRI Images. Researchers are still trying to modify the PDEs for specific applications [5]. Another important step in tumor detection is segmentation. Segmentation is the main way of distinguishing the objects from the background. For MRI brain image segmentation, various methods have been developed, namely, fuzzy set theory [6], mathematical morphology [7], clustering method [8], boundary based technique [9], region based method [10], and hybrid technique [11]. The choice of the specified method depends on the particular problem. Even though lots of methods for brain tumor segmentation have already been proposed, region based segmentation is a technique to determine the region directly whereas other methods achieve this goal by searching for the edges between regions based on breaches in gray levels or color properties [12–15]. Fig. 1 Overall workflow of the proposed methodology

J Med Syst (2013) 37:9938

In this paper, an improved image denoising and segmentation approach is proposed for detecting brain tumor automatically from MRI data sets. We have modified the PDEs by introducing a compass operator to have the benefit of preserving the edges from eight different angles. A new approach of skull stripping is also applied as a part of tumor detection process. An automatic region growing method based on single seed point selection is applied for detecting the brain tumor automatically. The rest of the paper is organized as follows. “Proposed methodology” contains the detailed stepwise methodology of the model. “Results and discussion” presents the experimental results and discussion. “Conclusion” concludes the paper.

Proposed methodology In our study, we propose a robust technique for automatic detection of tumor from brain MRI images. The proposed model consists of three major parts: Pre-processing, Segmentation, and Post-processing, which include noise removal and skull stripping in the preprocessing phase, seed selection and region growing technique in the segmentation part, and object labeling and image adjustment in the postprocessing phase. As mentioned above, we have modified the fourth order PDE in terms of edge preservation for denoising the MRI images. A new skull stripping technique using morphological operation has been used. The proposed method also includes a new concept of automatic seeded region growing technique. Figure 1 shows the overall workflow of the proposed model. Noise removal using modified fourth order PDE In case of edge preservation, the Laplacian operator is used in the fourth order PDE. Laplacian operator describes only the edge magnitude without edge orientation or directions. Moreover, it responds twice to some borders or edges in the

J Med Syst (2013) 37:9938

Page 3 of 10, 9938

image [5]. Preserving the orientation and directions of all anatomically significant edges while detecting the tumor, is a challenging task. Here, compass operator has been used instead of the Laplacian operator with an aim to preserve the edges from eight different angles. The compass edge detector is an appropriate way to estimate the magnitude and orientation of an edge. For each pixel, the local edge gradient magnitude is estimated by the compass operator with the maximum response of all eight kernels at this pixel location. The values −1, −2, +1, +2 in Fig. 2 are the coefficients of the kernels. By rotating the coefficients of the kernels, the different scale in the magnitude image can be achieved [16]. Compass operator is able to find the direction of the maximal growth of a function, for example, from black (f(i,j)=0) to white (f(i, j)=255). Gradient compass operator blended with the fourth order PDE [15] is shown below: @I =@t ¼ Ok ½C ðOk I ÞOk I 

ð1Þ

where the compass gradient is defined as I ðx; yÞ ¼ max IK ðx; yÞ K

ð2Þ

In Eq. (1), O is the operator and k is the direction of the compass of the image I. The suggested PDE model attempts to eliminate noise and conserve the edges by approximating an observed image with a piecewise planar image. Using the compass operator, the image is convolved with various kernels (eight directions, i.e., 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°). Figure 2 shows some examples of compass operator. The compass operator has the aptitude to divide an image window into half, thereafter compares the two halves to observe if they are not identical. Most of the edge detectors compute an average value for each half and also compute the Euclidean distances. The compass operator allows multiple values to exist on each side. Equation (1) is associated with the following functional expression: Z EðIÞ ¼ f ðjOk I jÞ@x@y ð3Þ

Planar images are the only global minimum of E(I) [5] for two dimensions if Z     ð5Þ E1 ðIÞ ¼ Ω jIxx j þ Iyy dxdy The cost function E(I) is convex under these circumstances [5] and the image is increasingly smoothed until it becomes a planar image. The image processed by the fourth order PDE will look comparatively less blocky than that characterized by other anisotropic dissemination or noise removal techniques. We have measured the oscillation of the noisy data using Eqs. (5) and (6). Z qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  2  2  2 E2 ðIÞ ¼ ð6Þ jIxx j2 þ Ixy  þ Iyx  þ Iyy  dxdy Ω

The main anomaly between Eqs. (5) and (6) is that, Eq. (6) is a rotational invariant for higher dimensions while Eq. (5) is not. For higher dimensional problems, E1(I) is simpler to implement. Therefore, by applying these processes (Eq. (6)), we can achieve a noise free image as presented in Fig. 3. Skull stripping Due to homogeneous intensities of the skull and brain tumor, it is a big problem to segment the tumor from the brain. Moreover, the skull is the non-cerebral tissue of the brain. Therefore, skull stripping is one of the major parts in brain MRI imaging. A background removal technique using morphological operation has been used here. In this study, the main focus of skull stripping is distortion of some cerebral tissues while removing background and later on restoration of foreground. Taking into consideration the skull as background of MRI image and the tumor as foreground, the morphological opening operation is applied and subtracted the resultant image from the original image to extract the background with some cerebral tissues. Now, after subtracting the image from the original one, the foreground can be achieved.

Ω

where Ω is the image support and Ok is the gradient compass operator. The global minimum of E(I) occurs when jOk I j ¼ 0 for allðx; yÞ 2 Ω

Fig. 2 Compass operator a 0° b 45°

ð4Þ

Fig. 3 a Input image b Image after denoising

9938, Page 4 of 10

J Med Syst (2013) 37:9938

For the first step, opening operation is applied, i.e., erosion, followed by dilation. Considering an image f(x) and the structuring function b(x), the grayscale dilation can be obtained as follows [12]: n o ðf  bÞðxÞ ¼ x : b bx \ f 6¼ f ð7Þ The erosion operation of image f by structuring element b can be obtained by [12], f Θ b ¼ fx : bx  f g

ð8Þ

Hence, Eq. (9) shows the opening operation [12], i.e., dilation of the erosion of image f(x) by structuring element b, as shown in Fig. 4: X ði; jÞ ¼ f  b ¼ ð f Θ bÞ  b

ð9Þ

where X(i,j) is the resultant image. It can also be written as S f b¼ bx which means that, it is the locus of translation bx f

of structuring element b inside the image f. The structuring element shown in Fig. 4 is used to probe the original images as opening operations. Each point in the structuring element may have a value. The grayscale images have the value ranges from 0 to 255, while for binary images, it is 0 and 1. Here, the elements are represented as “1” and ‘0’. Figure 4 shows the structuring element that we have exhaustively used for this operation of all images. Now from Eq. (10), we can get the skull of the brain images as below. Y ði; jÞ ¼ f ði; jÞ  X ði; jÞ

ð10Þ

Using 5×5 small structuring element, the skull is dilated lighter, which can be expressed in Eq. (11). It helps to remove the skull completely from the input image. Z ði; jÞ ¼ yði; jÞ  c

ð11Þ

Finally, the image at Fig. 5(b) is obtained from Eq. (12) as follows.

Fig. 5 a Image after detecting skull b Image after stripping

I ði; jÞ ¼ f ði; jÞ  Z ði; jÞ

ð12Þ

Figure 5(a) shows the detecting skull and Fig. 5(b) shows the image after skull stripping. Automated region growing segmentation Regions are identified for feature extraction and/or image recognition. After skull removal, the image is fed into an automatic segmentation algorithm that considers pixel homogeneity with the seed point selection. Instead of selecting a seed point manually [17], it is chosen automatically. The tumor region will grow according to the shape of the tumor once the seed point is automatically selected. Seed point selection is based on some criterion, for example, pixels in a certain gray-level range, pixels evenly spaced on a grid, etc. The initial region begins at the exact location of these seeds. Now, this initial region point (seed point) selection can be done manually by the user from the image, which is known as manual seed selection. Conversely, it can be done based on pixel intensity, gray level by the algorithm or method itself, which is called automatic seed point selection. The standard seed point must have three properties: (i) (ii)

The seed point must be in the region. The distance from the seed point to all of its neighboring pixels should be small enough to consecutive propagation. (iii) Most of the pixel in the ROI (Region of Interest) belongs to the region. Equation (13) is developed according to these three criteria and also by multiplying three weights as they have different quantities; we need to multiply three weights: w 1, w2 , and w3 to balance them into this function as below:

Fig. 4 Structuring element (disk)

f ðx; yÞ ¼ w1  g1 ðx; yÞ þ w2  g2 ðx; yÞ þ w3  g3 ðx; yÞ

ð13Þ

J Med Syst (2013) 37:9938

Page 5 of 10, 9938

Fig. 8 a 4-connectivity b 8-connectivity

Fig. 6 Specification of Euclidean distance, spatial distance, and ROI

The cost function is built by adding the three subfunctions corresponding to the three criteria as follows: (i)

magnitude. This approach is characterized by a cost function, which describes some features of the images around the seed. If there is any homogeneity then the pixel under consideration will be added. The gradient based cost function is [17],

Euclidean distance from the center point of ROI to the pixel. (ii) Spatial distance from the centroid to the pixel. (iii) The sum of the Euclidean distances from pixel to its neighbors.

Gn ¼

A suitable diagram is presented in Fig. 6 to specify the Euclidean distance, spatial distance, ROI, region, seed pixel, and neighbor pixel. Equation (13) is applied to each pixel in the ROI and the lowest value is chosen as our seed point. To select the ROI, we obtained a threshold value from the histogram of the image. We calculate the seed point from the gray image of 0–255 level intensities. The method for selecting ROI calculates the weight, mean, and variance from the histogram of the image. At every iteration, it takes each intensity level as the threshold and calculates within class variance. The minimum result of within class variance for the intensity level is the threshold point of the gray image. Then, the intensities are divided into two clusters: one is higher intensity area and another is lower. The higher intensity area is taken as ROI for the region growing procedure (a regionbased image segmentation approach). The main goal of this process is to partition an image into regions. We used the region growing based on pixel similarity using intensity values and their gradient direction and

Gm ¼

Fig. 7 Segmented tumor region

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi G2x þ G2y þ G2z KGmax

where 0 < Gn < 1

ð14Þ

Here, K is a constant for controlling the region growing and Gmax is the maximum gradient magnitude. Gmax  Gðx; yÞ where 0 < Gm < 1 Gmax  Gmin

ð15Þ

Here, G(x,y) is the gradient of the pixel under consideration. In our algorithm, we used 8-neighborhood for growing the region. Considering the seed point, the technique visits each neighbor pixel from the seed point. If the homogeneity criteria match with the seeded pixel then the pixel is added to the stack. Morphological operation Morphological operation has been applied after region growing to dilate the image so that the tumor can be visible clearly and sharply. This is also almost same like grayscale dilation as stated in Eq. (7) though it is on binary image. Equation (16) shows the dilation operation on binary image. We used a small 1×1 disk shape structuring element [12], required in binary morphological operation. Fig. 9 Labeled image

9938, Page 6 of 10

J Med Syst (2013) 37:9938

Fig. 10 Flowchart of the labeling technique

   I ði; jÞ  B ¼ z 2 EðBs Þz \ I 6¼ f

ð16Þ

where B is the structuring element, E is a Euclidean space or an integer grid, and Bs denotes the symmetric of B. We get the binary image of Fig. 7 after applying region growing technique.

Fig. 11 a Histogram of original image b Histogram equalized image

Object labeling In order to discover the number of objects, an essential observation is made in the neighborhood. The 4connectivity and 8-connectivity are the two types of neighborhood, used to label a binary image. Figure 8(a) shows the 4-connectivity and Fig. 8(b) shows the 8-connectivity

J Med Syst (2013) 37:9938

Page 7 of 10, 9938

Fig. 12 Tumor detected on final enhanced image

40 Proposed Exsisting

38

PSNR

36 34 32 30 28 26

neighborhood to label the object. Here, “1” represents the foreground pixel and “0” denotes the background pixel. The foreground pixels can be connected as neighborhood (not background pixel). In the 8-connectivity neighborhood, the pixels are connected horizontally, vertically, and diagonally. Whereas, in 4 connectivity, the pixels are connected only horizontally and vertically (not diagonally). Therefore, 8-connectivity neighborhood is used to label every pixel that touches one of their edges or corners. The 8-connectivity neighborhood is first found to label the target image. Afterwards, a) The input image is encoded by run length encoding. b) The runs are scanned and assigned initial labels, followed by tracing the label equivalences. c) The equivalence classes are then determined. d) Re-labeling is done on the run based equivalence classes. In order to color each object using different colors, the labeled image has been converted to RGB image. Figure 9 shows the final labeled image and Fig. 10 shows the workflow of the labeling technique.

Table 1 The PSNR calculation Image 1 2 3 4 5 6 7 8 9 10 Average

PSNR (Proposed)

PSNR (Existing)

36.51 36.44 37.49 35.89 36.77 39.26 36.45 35.63 34.78 35.74 36.49

27.31 27.29 27.43 27.91 27.54 27.12 27.69 26.71 26.64 27.21 27.28

1

2

3

4

5 6 Image

7

8

9

10

Fig. 13 Comparison between the proposed and the existing noise removal technique

Image adjustment and final output Labeling helps to locate the position of the tumor, although resulting image has a very low contrast to be observed by the physicians. A simple contrast adjustment technique has been applied to generate a clearly visible output, by raising the comprehensive contrast. In the histogram, the intensities are better distributed and we get higher contrast from the low contrast data [18]. As our final output is on the grayscale image, considering a grayscale image {G} where ni is the number of instances of gray level i, the probability of an instance of a point in the image can be computed by the following: pG ðiÞ ¼ pðG ¼ iÞ ¼

ni n

; 0iL

ð17Þ

where L is the total number of gray levels, n is the total number of pixels, and pG(i) is the histogram for pixel value i, which is normalized to [0,1]. The cumulative distribution function (also called accumulated normalized histogram) can be expressed by the following [18]: cdfG ðiÞ ¼

i X

pG ðjÞ

ð18Þ

j¼0

Table 2 Performance measure of the proposed method

Measured procedures

Result

TPF FPF PV TNF Accuracy

0.9899 0.03 0.98 0.98 99.46

9938, Page 8 of 10

J Med Syst (2013) 37:9938

Table 3 Comparison between the existing tumor segmentation algorithms and the proposed technique Method Harikrishna Rai G.N et al. [17] Liu et al. [23] Xiao Xuan, Qingmin Liao [24] Clark et al.[25] Prastawa et al. [26] Fletcher-Health et al.[27] Sahar Ghanavati et al. [28] M. Usman Akram et al. [29] Proposed Method

Accuracy 92.37 99 96.82 70 80 90 90.11 97 99.46

If we create a transformation in the form I = T(G) to generate a new image {I}, such that its cumulative distributed function will be linearized across the value for some constant k, i.e., cdfI ðiÞ ¼ ik

ð19Þ

The characteristics of the CDF allow us to perform such a transformation and it is defined as [18]: I ¼ T ðGÞ ¼ cdf ðGÞ

ð20Þ

In Eq. (20), within the range of [0,1], T maps the levels. The following function is employed on the obtained results of mapping the values back into their original range. I 0 ¼ I:ðmaxfGg  minfGgÞ þ minfGg

ð21Þ

Figure 11 shows the histogram of original image and histogram equalized image. Finally, we get the enhanced brain image with the clearly detected tumor region, as presented in Fig. 12.

Results and discussion The proposed technique has been tested on publicly available brain tumor images obtained from medical database (http://www.imageprocessingplace.com/root_ files_V3/image_databases.htm). These images contain 8-bit per channel. The performance evaluation of the proposed technique as well as comparison is carried out based on the two aspects. One is the Peak Signal to Noise Ratio (PSNR) pertaining to noise removal and another is the tumor detection accuracy of brain MRI images, which is verified by the physician. The PSNR calculation has been carried out on 10 noisy brain MRI images of different resolutions ranging between 252×

Fig. 14 a Original image b Brain tumor detected image

289 and 383×500. The PSNR values we achieved are better than the existing PDE based algorithm [19]. The PSNR values between the original image and the noise free image are found using Eq. (23). The PSNR can be

J Med Syst (2013) 37:9938

Page 9 of 10, 9938

easily defined by calculating MSE (Mean Square Error) using Eq. (22). MSE ¼

m1 X n1 1 X ½I ði; jÞ  k ði; jÞ2 mn i¼0 j¼0

ð22Þ

where I(i,j) is noise-free m × n monochrome image and K is the noisy image.   MAX2 PSNR ¼ 10log10 MSEI   ð23Þ ffiffiffiffiffiffiffiI ¼ 20log10 pMAX

pixel, TN (true negative) is the correctly rejected object/ pixel, and FN (false negative) is the incorrectly rejected object/pixel. For mathematical elaboration, we may refer to [20–22]. Table 2 depicts the results obtained by using the proposed method, while Table 3 compares the results obtained from the proposed technique with those obtained by other known or existing tumor segmentation techniques [17, 23–30]. Besides, Fig. 14 illustrates the original image and segmented brain tumor image.

MSE

¼ 20log10 ðMAXI Þ  10log10 ðMSEÞ

Conclusion

Here, MAXI is the maximum possible pixel value of the image where the pixels are represented using 8 bits per sample, that is 255. Table 1 shows the PSNR of 10 different images, where the obtained average PSNR is 36.49 and Fig. 13 shows the comparison, which confirms that the proposed noise removal technique is quite powerful and well suitable in our study. Gold standard image is used in all experimental settings. The performance evaluation and comparisons are carried out based on the four measurement procedures, namely, True Positive Fraction (TPF), False Positive Fraction (FPF), Specificity (TNF), and Predictive Analysis (PV) [20]. The accuracy is computed by the ratio of the total number of correctly classified points to the number of points in the image [21]. TPF represents the fraction of pixels accurately detected as tumor pixels, which is also known as sensitivity, whereas FPF represents the fraction of the pixels erroneously detected as tumor pixels. Specificity measures the negative that are correctly identified (the fraction of pixels correctly identified as not in the tumor pixel). The PV is the probability that a pixel which has been classified as bright lesion, is really a bright object. More specifically, the measurement procedures can be expressed as follows: TPF ¼

TP TP þ FN

ð24Þ

TNF ¼

TN TN þ FP

ð25Þ

FPF ¼

FP FP þ TN

ð26Þ

PV ¼

TP TP þ FP

ð27Þ

where TP (true positive) is the correctly identified object/ pixel, FP (false positive) is the incorrectly identified object/

In this study, a robust technique for brain tumor detection is presented by modifying the fourth order PDE for denoising the MRI brain images. The concepts of skull removal as well as seeded region growing segmentation are included to detect the tumor efficiently. We have tested our proposed method on the MRI image database and it offers good PSNR results (with respect to noise removal) and 99.46 % accuracy in terms of tumor detection. However, in future, more advanced multiple seed selection technique can be applied to classify the brain tumors using SVM (Support Vector Machine) or Neural Network classifier from the 3D MRI brain images.

Conflict of interest The authors declare that they have no conflict of interest.

References 1. Shen, S., Sandham, W., Granat, M., and Sterr, A., MRI fuzzy segmentation of brain tissue using neighborhood attraction with neural network optimization. IEEE Trans. Inf. Technol. Biomed. 9(3):459–467, 2005. 2. Diaz, I., Boulanger, P., Griener, R., and Murtha A., A critical review of the effects of denoising algorithms on MRI brain tumor segmentation. 33rd Annual International Conference of the IEEE EMBS, Boston, Massachusetts, USA, pp. 3934–3937, 2011. 3. Prashanta, H. S., Shahidhara, H. L., Murthy, K. N. B., and Madhavi, L. G., Medical image segmentation. Int. J. Comput. Sci. Eng. 2(4):1209–1218, 2010. 4. Krajsek, K., and Mester, R., The edge preserving winner filter for scalar and tensor valued image. DAGIM, pp. 91–100, 2006. 5. Lysaker, M., Lundervold, A., and Tai, X.-C., Noise removal using fourth order partial differential equation with applications to medical magnetic resonance images in space and time. IEEE Trans. Image. Proc. 12(12):1579–1590, 2003. 6. Leung, C. C., Chen, W. F., Kwok, P. C. K., and Chan, F. H. Y., Brain tumor boundary detection in MR image with generalized fuzzy operator. Proceedings of the 2003 International Conference on Image Processing. Vol. 2, pp. II - 1057–60, Sept. 2003.

9938, Page 10 of 10 7. Rajeswari, R., and Anandhakumar, P., Segmentation and identification of brain tumor MRI image with Radix4 FFT techniques. Eur. J. Sci. Res. 52(1):100–109, 2011. 8. Koley, S., and Majumder, A., Brain MRI segmentation for tumor detection using cohesion based self merging algorithm. 2011 IEEE 3rd International Conference on Communication Software and Networks (ICCSN), pp. 781–785, May 2011. 9. Zhu, Y., and Yan, H., Computerized tumor boundary detection using a Hopfield neural network. IEEE Trans. Med. Imaging 16(1):55–67, 1997. 10. Prastawa, M., Bullitt, E., Ho, S., and Gerig, G., A brain tumor segmentation framework based on outlier detection. Med. Image. Anal. 8(3):275–283, 2004. 11. Kharrat, A., Gasmi, K., Messaoud, M. B., Benamrane, N., and Abid, M., A hybrid approach for automatic classification of brain MRI using genetic algorithm and support vector machine. Leonardo J. Sci. 9(17):71–82, 2004. 12. Gonzalez, R. C., and Woods, R. E., Digital image processing. Prentice Hall, Upper Saddle River, NJ, 2004. 13. You, Y.-L., Xu, W., Tannenbaum, A., and Kaveh, M., Behavioral analysis of anisotropic diffusion in image processing. IEEE Trans. Image. Proc. 5(11):1539–1553, 1996. 14. Kaus, M. R., Warfield, S. K., Nabavi, A., Black, P. M., Jolsez, F. A., and Kikinis, R., Automated segmentation of MR images of brain tumors. Radiology 218:586–591, 2001. 15. Lysaker, M., Osher, S., and Tai, X.-C., Noise removal using smoothed normals and surface fitting. IEEE Trans. Image Process. 13(10):1345–1357, 2004. 16. Abdou, I. E., and Pratt, W. K., Quantitative design and evaluation of enhancement/thresholding edge detectors. Proc. IEEE 67(5):753–763, 1979. 17. Rai, G. N. H., and Nair, T. R. G., Gradient based seeded region grow method for CT angiographic image segmentation. InterJRI Comput. Sci. Netw. 1(1):1–6, 2009. 18. Beghdadi, A., and Negrate, A. L., Contrast enhancement technique based on the local detection of edges. Comp. Vision Graph. Image Process. 46(2):162–174, 1989. 19. Chen, B., Cai, J.-L., Chen, W.-S., and Li, Y., A multiplicative noise removal approach based on partial differential equation model. Mathematical Problems in Engineering. Vol. 2012, Article ID: 242043, March 2012 (Accepted).

J Med Syst (2013) 37:9938 20. Reza, A. W., and Eswaran, C., A decision support system for automatic screening of non-proliferative diabetic retinopathy. J. Med. Syst. 35(1):17–24, 2011. 21. Reza, A. W., Eswaran, C., and Hati, S., Automatic tracing of optic disc and exudates from color fundus images using fixed and variable thresholds. J. Med. Syst. 33:73–80, 2009. 22. Reza, A. W., Eswaran, C., and Dimyati, K., Diagnosis of diabetic retinopathy: automatic extraction of optic disc and exudates from retinal images using marker-controlled watershed transformation. J. Med. Syst. 35:1491–1501, 2011. 23. Liu, J., Udupa, J. K., Odhner, D., Hackney, D., and Moonis, G., A system for brain tumor volume estimation via MR imaging and fuzzy connectedness. Comput. Med. Imaging Graph. 29(no. 1):21– 34, 2005 [22]. 24. Xuan, X., and Liao, Q., Statistical structure analysis in MRI brain tumor segmentation. Fourth International Conference on Image and Graphics, pp. 421–426, August 2007. 25. Clark, M. C., Hall, L. O., Goldgof, D. B., Velthuizen, R., Murtagh, R., and Silbiger, M. S., Automatic tumor segmentation using knowledge-based techniques. IEEE Trans. Med. Imaging 17(2):187–201, 1998. 26. Prastawa, M., Bullitt, E., Ho, S., and Gerig, G., A brain tumor segmentation framework based on outlier detection. Med. Image Anal. J. 8(3):275–283, 2004. 27. Fletcher-Heath, L. M., Hall, L. O., Goldgof, D. B., and Reed Murtagh, F., Automatic segmentation of nonenhancing brain tumors in magnetic resonance images. Artif. Intell. Med. 21(1–3):43– 63, 2001. 28. Ghanavati, S., Junning, Li, Liu, T., Babyn, P. S., Doda, W., and Lampropoulos, G., Automatic brain tumor detection in magnetic resonance images. 9th IEEE International Symposium on Biomedical Imaging, pp. 574–577, May 2012 . 29. Akram, M. U., and Usman, A., Computer aided system for brain tumor detection and segmentation. 2011 International Conference on Computer Networks and Information Technology, pp. 299–302, July 2011. 30. Faisal, A., Parveen, S., Badsha, S., and Sarwar, H., An improved image denoising and segmentation approach for detecting tumor from 2-D MRI brian images. International Conference on Advanced Computer Science Applications and Technologies, Kuala Lumpur, November 2012.