IEEE 2 Column Format

1 downloads 0 Views 606KB Size Report
V.R.Siddhartha Engineering College ... evaluation of large amounts of data, MRI-based medical image ... The aim of medical image segmentation in brain.
2015 International Conference on Advanced Computing and Communication Systems (ICACCS -2015), Jan. 05 – 07, 2015, Coimbatore, INDIA

An Automated Segmentation of Brain MRI for detection of Normal Tissues using Improved Machine Learning Approach M.Y.Bhanumurthy

Koteswararao Anne

Dept. of ECE Vasireddy Venkatadri Institute of Technology Guntur – 522508. A.P, INDIA [email protected]

Dean, Academics V.R.Siddhartha Engineering College Kanuru, Vijayawada, A.P, INDIA [email protected]

Abstract— Due to an increased need for efficient and objective evaluation of large amounts of data, MRI-based medical image analysis is gaining attention in recent times. The goal is to simplify an image into something that is more meaningful and making it easier to analyze. The aim of medical image segmentation in brain MRI is to separate the region of interest from the background after denoising and skull removal. Accurate segmentation of normal and abnormal tissues is still a challenge for researchers. In this paper, we propose a fully automated segmentation of normal tissues viz., white matter (WM), gray matter (GM) and cerebro spinal fluid (CSF) from brain MRI using an improved machine learning approach that uses Neuro-fuzzy as classifier. The segmentation is carried out using gradient method and orthogonal polynomial transform. The performance of our method is assessed with metrics such as false positive rate (FPR), false negative rate (FNR), specificity, sensitivity and accuracy. Also, the entire procedure is developed as a graphical user interface (GUI) which results in automated classification and segmentation. Keywords—Machine Learning; Neuro-Fuzzy classifier; Segmentation; Normal Tissues; Gradient Method; Orthogonal Polynomial Transform; GUI

I.

INTRODUCTION

The brain is the frontal most part of the central nervous system. It forms the central nervous system (CNS) along with the spinal cord. The cranium, a bony box in the skull guards it. Because of our brain we do lot of things like, to think, act, reason, walk, talk, the list is never-ending. By means of Magnetic Resonance Imaging (MRI), doctors and researchers can examine noninvasively the structure and function of the brain. In fact, the MRI image is a thin horizontal slice of the brain. Segmentation is a significant implementation in medical image processing and it has been helpful in several applications. It may be helpful to categorize image pixels into anatomical regions in some applications such as bones, muscles, and blood vessels, while in others into pathological regions, such as cancer, tissue deformities, and multiple sclerosis lesions. To detach an image into regions that are homogeneous with respect to one or more characteristics is the primary objective in segmentation process [1]. The aim is to separate the whole image exactly into sub regions included gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF) spaces of the

[978-1-4799-6438-3/15/$31.00 ©2015 IEEE]

brain [2] in MR images processing. As in a number of neurological disorders like multiple sclerosis (MS) and Alzheimer’s disease, the volume changes in total brain, WM, and GM can give major notification about neuronal and axonal loss [3]. The rest of the paper is organized as follows: After this introduction, the next section surveys several works that were proposed for segmenting the tissues in the brain MRI. Section III illustrates and explains about our proposed technique of automated classification and segmentation of the normal tissues in brain MRI images utilizing Neuro-fuzzy classifier. The outcomes regarding the performance of our proposed work are described in Section IV and our paper is summed up with the conclusion in Section V. II. LITERATURE SURVEY Various algorithms have been proposed for brain MRI segmentation [4-8] like watershed algorithm, eSnake algorithm and genetic algorithm which are based on the homogeneity of image. Actually, we have to solve the problem with new technique as the intensity in homogeneity is impact on every image. The successful techniques were thresholding [8], regiongrowing [9] and clustering [1]. Wells [9] buildups a latest statistical approach based on the expectation-maximization (EM) algorithm, however the results are too reliant on the initial values, very time consuming and were just looking for local maximum point. Arnaldo Mayer and Hayit Greenspan [10] have offered an automated segmentation framework for brain MRI volumes based on adaptive mean-shift grouping in the joint spatial and intensity feature space. By utilizing a subject-specific tissue probabilistic atlas produced from longitudinal data, Feng Shi et al. [11] have offered a framework for presenting neonatal brain tissue segmentation. Mert R. Sabuncu et al. [12] have examined a generative model that guides to label fusion style image segmentation techniques. Dalila Cherifi et al. [13] have illustrated normal tissue’s recognition than tumor extraction (applied for GBM and MS diseases). Juin-Der Lee et al. [14] have offered the most statistical segmentation methods in the literature and have presumed that either the intensity allocation

2015 International Conference on Advanced Computing and Communication Systems (ICACCS -2015), Jan. 05 – 07, 2015, Coimbatore, INDIA

of every tissue variety was Gaussian, or the logarithmic transformation of the raw intensity was Gaussian. III. PROPOSED APPROACH

Extracted Features

Energy Entropy Homogeneity Contrast Correlation

Entropy Ep, Homogeneity Hm, Contrast Cn, and Correlation Cr are the features extracted from the brain MRI images [15]. These features are then applied to a Neuro-fuzzy classifier that classifies the images into normal and abnormal. Our improved machine learning approach is a Neuro-fuzzy system that has a three-layered architectural design. Our classifier is a fuzzy based system that is trained by a learning algorithm derived from neural networks. The learning algorithm performs only on the local information and provides the local modifications in the fuzzy system. In general, our Neuro-fuzzy system generates very powerful solutions when compared to the use of system components individually. A. Segmentation of Normal Tissues Before segmenting the normal tissues, a procedure called pre-processing is performed on the normal images. The brain MRI images that are classified as normal are first transformed into gray scale images in order to apply the morphological operations [16] on the images. Using region based binary mask extraction procedure, the brain cortex is stripped. Generally, the brain cortex is observed as a ring around the brain tissues. The Skull Stripping (SS) technique removes the ring that surrounds the brain tissues and the normal image attained after the preprocessing is denoted as I SS . From the pre-processed normal images, normal tissues like white matter (WM) and gray matter (GM) are segmented using gradient method and cerebro spinal fluid (CSF) is segmented using orthogonal polynomial transform.



Normal Images

Abnormal Images

B. Segmentation of WM and GM For segmenting the white and gray Matters, the preprocessed skull stripped image I SS is subjected to gradient technique. The gaussian convolution filter utilized in this technique makes the image I SS into a smoothed image I S which is then processed with gradient operation. The gradient of two variables x and y are specified as follows:

Normal Image Tissue Segmentation

 I S ( x, y )  White Matter

Gray Matter CerebroSpinalFluid

Fig.1. Proposed Methodology [15]

Fig. 1 shows our proposed methodology for segmentation of normal tissues using improved machine learning approach. In our work, we have used Brainweb datasets for experimentation which includes both normal and abnormal images. Energy Eg,

I S  I S  e f x y



The gradient values are useful to mark the current edges in the image that are specified in the following equations (2) and (3).

S  x( e ) 2  y ( f ) 2

(2)

1 (1  S )

(3)

EM 

2015 International Conference on Advanced Computing and Communication Systems (ICACCS -2015), Jan. 05 – 07, 2015, Coimbatore, INDIA

The process of binarization is then carried out on the edge marked ( EM ) image wherein the value of gray level of every pixel is estimated by means of a global threshold value T g . The resultant image after the binarization process is denoted as I B . Next, by the use of morphological opening and closing operation, the small holes and small objects from the image I B are removed. In our work, the WM and GM tissues were segmented by means of the intensity values as given in equation (4)

categorization correctness outcome. Table III illustrates the segmentation results of normal tissues. TABLE I:

DESCRIPTION OF TP, TN, FP AND FN VALUES

Description Actually normal image Actually abnormal image

Classified as normal image

Classified as abnormal image

TP

FN

FP

TN

TABLE II:

EFFECTIVENESS OF USING NEURO-FUZZY CLASSIFIER

 WM , if I Bi 1 IWG    GM , if I Bi  0

Metrics TP TN FP FN FPR FNR Sensitivity Specificity Accuracy

(4)

If the intensity value of the binarized image is 1, then the tissue is considered as WM and subsequently segmented. The tissue is considered as GM and subsequently segmented, if the intensity value is 0. C. Segmentation of CSF CSF tissue from the image

CLASSIFICATION

RESULTS

Values 8 14 0 1 0 6.667 88.9% 100% 95.65%

TABLE III: SEGMENTATION OUTCOMES OF WM, GM AND CSF

I SS is segmented by orthogonal

polynomial transform (OPT) according to the equation given below:

2

 I SS 3  I CSF  Sin  ( i )   0.05 * rand ( I SS )   100   

(5)

IV. RESULTS AND DISCUSSION The entire process of feature extraction, classification, preprocessing and segmentation of normal tissues is developed as a GUI system. The results for the automated detection and segmentation of normal tissues of brain MRI are shown in figs. 3-8. The performance of our classifier is assessed using the metrics false positive ratio (FPR), false negative ratio (FNR), specificity, sensitivity and accuracy. The basic count values such as true positive (TP), true negative (TN), false positive (FN) and false negative (FN) are used by these metrics. Table I describes the count values for classifying the images into normal and abnormal and table II explains the categorization efficiency outcomes for the normal and abnormal images with various metric values. FPR and FNR of values 0 and 6.667 indicate that our proposed classifier classifies the images into normal and abnormal with very low error rates. An accuracy of 95.65% indicates an improvement in the

Im ag e No . 1 2 3 4 5

WM

GM

CSF

Speci ficity (%)

Sensi tivity (%)

Accu racy (%)

Speci ficity (%)

Sensi tivity (%)

Accu racy (%)

Speci ficity (%)

Sensi tivity (%)

Accu racy (%)

96.9 91.8 97.9 96.8 89.0

11.8 29.2 8.09 15.3 39.2

95 91 95 94 86

90.6 94.4 91.0 95.4 94.5

54.7 28.5 85.2 44.5 51.3

84 84 90 84 85

98.3 98.2 99.2 98.4 99.5

4.52 7.31 1.73 5.05 1.59

92 88 95 95 99

Overall, our proposed technique gives an accuracy of 92.2%, 85.3% and 93.8% for WM, GM and CSF segmentations respectively on average. V. CONCLUSION An automated and improved machine learning approach that uses Neuro-fuzzy as classifier to detect and segment normal tissues like WM, GM and CSF in brain MRI images is proposed in this paper. The extracted features from the images were applied to our classifier which classifies the images into normal and abnormal. The normal images were pre-processed first and then segmented efficiently by our proposed method. The testing was performed with the Brainweb images dataset. Our method is evaluated using the performance measures FPR, FNR, sensitivity, specificity and accuracy. The efficiency of the classification of images is very high which is evident from the accuracy outcomes and the segmentation of normal tissues also offers very accurate outcomes. From the results, we have showed that the Neuro-fuzzy classifier utilized in our proposed work achieves a very good accuracy of 95.65% in categorizing

2015 International Conference on Advanced Computing and Communication Systems (ICACCS -2015), Jan. 05 – 07, 2015, Coimbatore, INDIA

the images into normal and abnormal. Also, the developed GUI system simplifies our entire process of loading and viewing the images and obtaining the required results. REFERENCES [1]

[2]

[3]

[4]

[5] [6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

J. C. Bezdek, L.O. Hall, L. P. Clarke, “Review of MR image segmentation techniques using pattern recognition,” Med. Phys., vol. 20, No. 4, pp. 1033-1048, 1993. K. Lim, A. Pfefferbaum, “Segmentation of MR brain images into cerebrospinal fluid spaces, white and gray matter,” J. Comput. Assist. Tomogr, vol. 13, pp. 588-593, 1989. Y.Zhang, M. Brady, S.Smith, “Segmentation of brain MR images through a hidden Markov random field model and expectation-maximization algorithm,” IEEE Trans Med. Imag., pp. 45–57, 2001. H Lv., K. H. Yuan, S. L. Bao, “An eSnake model for medical imaging segmentation”, Progress in Natural Science, vol. 15, No. 5, pp. 424-429, 2005. D. L. Pham, J. L. Prince, “Adaptive fuzzy segmentation of magnetic resonance images,” IEEE Trans. Med. Imag., pp. 737-752, 1999. A. F. Goldszal, C. Davatzikos, D. L. Pham, M. X. H. Yan, et al, “An image processing system for qualitative and quantitative volumetric analysis of brain images,” J. Comput. Assist. Tomogra, vol. 22, No. 5, pp. 827-837, 1998. James B. Arnold et al., “Qualitative and quantitative evaluation of six algorithms for correcting intensity nonuniformity effects”, NeuroImage, 13, pp. 931-943, 2001. R. Moller, R. Zeipelt. “Automatic segmentation of 3D-MRI data using a genetic algorithm”, IEEE International Workshop on Medical Imaging and Augmented Reality, pp. 278 – 281, 2001. W.M.Wells III, W. E. L. Grimson, R. Kikinis. “Adaptive segmentation of MRI data”, IEEE Trans. Medical Imaging, vol. 15, No.4, pp. 429-442, 1996. Arnaldo Mayer , Hayit Greenspan, “An Adaptive Mean-Shift Framework for MRI Brain Segmentation”, IEEE Transactions On Medical Imaging, Vol. 28, No. 8, August 2009. Feng Shi, Yong Fan, Songyuan Tang, John H. Gilmore, Weili Lin, Dinggang Shen, “Neonatal brain image segmentation in longitudinal MRI studies”, NeuroImage, 49, pp. 391-400, 2010. Mert R. Sabuncu, B.T. Thomas Yeo, Koen Van Leemput, Bruce Fischl and Polina Golland, “A generative model for image segmentation based on label fusion”, IEEE Transactions on Medical Imaging, 2009. Dalila Cherifi, M.Zinelabidine Doghmane, Amine Nait-Ali , Zakia Aici,Salim Bouzelha, “Abnormal tissues extraction in MRI brain medical images”, IEEE WOSSPA, pp. 357-360, 2011. Juin-Der Lee, Hong-Ren Su, Philip E. Cheng, Michelle Liou, John A. D. Aston, Arthur C. Tsai, and Cheng-Yu Chen, “MR image segmentation using a power transformation approach”, IEEE Transactions On Medical Imaging, Vol. 28, No. 6, pp. 894-905, 2009. M.Y.Bhanumurthy, Koteswararao Anne, “An efficient Neuro-fuzzy based segmentation of normal tissues in brain MRI (BMRI) using Extensive Feature Set”, Journal of Biomedical Engineering and Medical Imaging, vol. 1, No. 5, pp. 1-22, 2014. Soumya Maitra,"Morphological edge detection using bit-plane decomposition in gray scale images", Proceedings of INDIACom, 2011.

2015 International Conference on Advanced Computing and Communication Systems (ICACCS -2015), Jan. 05 – 07, 2015, Coimbatore, INDIA

Fig.3. GUI Screenshot

Fig.6. Result for Image 3

Fig.4. Result for Image 1

Fig.7. Result for Image 4

Fig.5. Result for Image 2

Fig.8. Result for Image 5