SEGMENTATION AND DETECTION OF ABNORMALITIES IN BRAIN MR IMAGES

Abstract

Magnetic Resonance Imaging (MRI) technique is a non-invasive diagnostic tool which plays a crucial role in the main stream of anatomical studies and pathological planning of brain abnormalities when compared to other available techniques like Computed Tomography (CT), because of its excellent soft tissue contrast and multi-planer acquisition capabilities. It is the best neuro-imaging tool for in vivo projections of structural aberrations like diffused brain differences, structural alterations etc. It is used for visualization and quantitative analysis of the area inside the tissue. Brain Hemorrhage, Brain Infarction and Brain Tumor are the most common brain deformities, leading to high morbidity and mortality worldwide. Albeit, multitude MRI pulse sequences are there (like T1, T2, proton density, FLAIR, DWI, ADC, FAT saturation etc.) each of them keeping their unique characteristic to discern tissue composition but the basic MR sequences namely T1-weighted, T2-weighted and DWI furnish relevant texture and intensity information of Hemorrhage, Infarct and Tumors. The lesion may have homogeneous or heterogeneous texture, both having different signal characteristics i.e. hypointensity, hyperintensity or isointensity. Lesions are characterized in accordance to their location and aetiology (underlying causes). The T1 and T2 characteristics of the phases of hemorrhage and infarction namely Early Hyperacute, Late Hyperacute, Acute, Subacute and Chronic are different depending on the aging of hematoma with time. Also, the degree of enhancement of tumors for the same class and for different class of tumor depends on MR acquisition and its morphology. Any diminutive change resulting from any kind of disease or insult can cause serious further implications. Compression or dilation from the normal anatomy due to accumulation of blood or due to necrosis by the disruption of atrioventricular blood flow, leads to cerebrovascular problems. So, the pulse sequences (T1-weighted, T2-weighted, DWI etc.) are considered in order to get information about parenchymal issues such as: obliteration of the gyri-sulci pattern, asymmetry, abnormal distinction between gray-white matter, hyper/hypo dense abnormalities in brain parenchyma, hydrocephalus or intraventricular blood in ventricular system, extra cranial soft tissue swelling, fracture, normal air content of the sinuses and the mastoid in bone. Segmentation of abnormalities in brain MR images is a crucial step for surgical and treatment planning. There are various methods reported in literature for segmentation and classification of brain lesions on brain MR image. Segmentation procedure focused either on similarity criterion like thresholding, watershed, Region Growing, and region splitting and merging or on discontinuity criterion like edge-based methods. In another kind of spatially guided approach where spatial relationship between pixels is used for segmentation are region [ii] (like region growing, split and merge, hybrid growing merging) and energy-based techniques (like active contour, graph based, watershed, marker-based watershed). The level set formulation of the active contour model is further categorized as region-based approach and edge-based approach. Besides edge-based active contour models, region-based models try to detect every region of interest (ROI) in the given image by combining the region-based information into their energy functional. In literature, these models have been found to be superior than the edge-based active contour models because of their unrestricted position of the initial contour and automatic detection of estimated boundaries. Also, these models have shown effective reasonable segmentation because of global energy minimization based on the statistical properties inside and outside the evolving contour which provide driving force during the evolution of the deformable shapes and keeps the regularity of the active contour. The methods reported in literature namely Chan- Vese (CV), Region Scalable Fitting (RSF) and Local Gaussian Distribution Fitting (LGDF) are region-based active contour method. The CV model provides better performance than the previous models because of its nature to acquire a large convergence rate. Though, image segmentation using the CV model depends on the placement of initial contour, Region Scalable Fitting (RSF) energy function obtains information relating to intensity of the local regions at an established scale in order to estimate the two fitting functions and forces the curve towards the calculated contour of the object. The regularity term available in this model is used to control the length of the object boundaries and prevents the over-segmentation. In LGDF model, the local image intensities are described by gaussian distributions with different means and variances where circular initial level set function is used which then evolve to the object boundary. Above discussed methods are sensitive to initialization of contour and noise. Image features are useful in identification, representation and description purpose. Several researchers in past have given the methods of texture feature extraction from various medical image datasets. Intensity Based Features (IBF), Gray Level Difference Matrix(GLDM), Laplacian of Gaussian Features (LoG), Rotation Invariant Circular Gabor Features (RICGF), Rotation Invariant Local Binary Patterns (RILBP), Gray Level Difference Statistics (GLDS), Neighbourhood Gray Tone Difference Matrix (NGTDM), Laws Textures Energy Measure (Laws TEM), Fractal Dimension Texture Analysis (FDTA), Statistical Feature Matrix (SFM), Fourier Power Spectrum (FPS), Local Binary Pattern (LBP) and Gaussian Pyramid based Local Binary Pattern (GPLBP) are mostly used feature extraction techniques applied in classification of input dataset into two class or multiclass objects. [iii] As dimensionality of the feature set increases, the amount of data needed to give reliable analysis grows exponentially. Feature selection is essentially helpful in dealing over-fitting problem and in reducing the training time. Many methods are reported in literature for feature reduction for instance Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Minimum Redundancy Maximum Relevance (mRMR) and Locally Linear Embedding (LLE). These are used to select most relevant attribute to form an optimal subset from the original feature pool. Final diagnosis of lesion is done through classification models. There are many methods available in the literature for detection of normal and abnormal brain as well as for multiclass classification of lesions using MR images. It has been recognized and emphasized time and again that early and correct classification of pathology can reduce casualty in emergency conditions. The multiclass classification studies reported in literature are based on Support Vector Machine (SVM), Deep Neural Network (DNN), Artificial Neural Networks (ANN), or hybrid methods using Genetic Algorithms (GA) like GA-ANN, GA-SVM. Although MR images of brain furnishes evident visualization of the brain anatomy, the insight picture details such as detection of lesion is not very clear or definite. Consequent problems arising during the identification of lesion on MR images are: artifacts developed due to change in its orientation relative to patient posture, inadequate segmentation of homogeneous, heterogeneous, diffused lesion bearing weak or false edges and classification of lesion; have to be dealt. For that purpose, this research work is basically directed to develop a CAD system constituting segmentation and classification algorithms for identification of abnormality on brain MR images, thereby, aiding the radiologist in decision making process. In this research work two major objectives have been considered: The first objective of this research is to develop a method for segmentation of MR images to extract region of interest (ROI). The second objective is to develop a classification scheme based on artificial intelligence techniques for diagnosing critical MR brain abnormalities like Brain Hemorrhage, Infarct and Tumor, which are leading causes of rising incidence in mortality among adults and children. In order to achieve the first objective of the present work, Modified Region Based Active Contour (MRBAC) is proposed to delineate region of interest in brain MR images. The proposed method utilizes the advantages of Local Gaussian Distribution Fitting (LGDF) energy model along with coping up with the difficulties handled by the method. The performance of conventional active contour method relies on factors such as: appropriate and accurate contour [iv] initialization, presence of intensity inhomogeneities and weak boundaries, sensitivity to noise and optimal configuration of contour propagation controlling parameters. Although LGDF method can discern well between two similar intensity regions; however, if the contour initialization is not accurate enough it can introduce many local minima resulting in improper segmentation. The proposed method is based on the LGDF method with the variation at the initialization step so that it does not get trapped at local minimum. Here, contour of specified shape (circular or rectangular) has not been taken but of optimal shape which is user defined has been used and it is said to be as zero level set. This reduces the number of iterations resulting in reduction of computational cost of the system. The developed MRBAC method is tested on 118 images out of which 34 images are taken of brain hemorrhage, 27 images are taken of brain infarct and 57 images are taken of brain tumor. Different performance metrics such as True Positive, False Positive, Accuracy, Jaccard Similarity Index, Dice Coefficient and Hausdorff Distance are computed for quantitative analysis. Accuracy is the proportion of pixel contained within the correctly extracted regions obtained by the test methods out of all the pixels of manually delineated region. DC and JSI are the similarity matrices and their values should be higher. HD values are used to measure the structural difference between the two objects and its value should be low signifying better segmentation. It has been observed that the result of the proposed segmentation provides higher segmentation accuracy in all the three abnormalities when compared with other state of the art methods like Region Growing and Distance Regularized Level Set Evolution (DRLSE) segmentation methods. It also provides better values of other averaged performance measures such as JSI, DC and HD. In order to achieve the second objective of the present work, Pattern Recognition tool in Neural Network (NN) Toolbox (of MATLAB®) has been used for the classification of selected anomalies from MRI image of the brain. In this work, two different classifiers named as Fusion Network are proposed: first one is for two class classification of data into normal and pathological brain and second one is for multiclass classification of data into Normal, Tumor, Hemorrhage and Infarct brain abnormalities in MR images. For first classifier model, input samples of 488 images (230 normal brain, 258 pathological brain MRI) are taken. To classify these samples into normal and pathological brain, total 1861 features have been calculated. These features include 5 First Order Statistical features, 26 Haralick Spatial Gray Level Dependence Matrices (SGLDM) features, 4 Gray Level Difference Statistics (GLDS), 5 Neighbourhood Gray Tone Difference Matrix (NGTDM), 4 Statistical Feature Matrix (SFM), 6 Laws Texture Energy Measures (TEM), 4 Fractal Dimension [v] Texture Analysis (FDTA), 2 Fourier Power Spectrum (FPS), 361 Local Binary Pattern and its variants (LBPu2, LBPri, LBPriu2) and 1444 Gaussian Pyramid based Local Binary Pattern with its variants (LBPu2, LBPri, LBPriu2). Dimensionality of these features is reduced to 500 from 1861 by using Minimum Redundancy Maximum Relevance (mRMR) method which assigns rank to each feature according to its relevance in classification between different class. Original dataset of 488 images is divided into two parts one is training data (TR data) and other is testing data (TE data) in the ratio of 70% to 30% respectively. After normalization, feature set from training data is presented to ten individual two-layer feedforward neural network with 6 neurons in hidden layer having sigmoidal activation function and 2 neurons in output layer having SoftMax activation function. Same feature set is presented to proposed Fusion Network classier which is an ensemble of all ten individual neural networks. The network is trained with scaled conjugate gradient backpropagation method. Individual as well as proposed Fusion Network Classifier is tested using 30 % of original data (TE data). The accuracy of proposed Fusion Network classifier comes out to be 94.5% for classification of normal and pathological brain. For second classifier model, input samples of 510 images (176 Normal brain, 123 Tumor brain, 107 Hemorrhagic brain and 104 Infarct brain MRI) are taken. To classify these samples into Normal, Hemorrhage, Infarct and Tumor brain, total 1861 features have been extracted from all classes of brain MR images. Using mRMR method, these features are then reduced to 500 most relevant feature set which have important role in discriminating input data between all four class. All 510 images of dataset are partitioned in to training data (TR data) and testing data (TE data) in the same ratio as previous i.e. 70% to 30%. Feature set obtained from training data is presented to ten separate two-layer feedforward neural network with 6 neurons in hidden layer having sigmoidal activation function and 4 neurons in output layer having SoftMax activation function. Same feature set is presented to proposed Fusion Network classier which is a combination of all ten individual neural networks. The network is trained with scaled conjugate gradient backpropagation method. All separate neural networks as well as proposed Fusion Network Classifier is tested using 30 % of TE data. The accuracy of proposed Fusion Network classifier comes out to be 92.8% for multiclass classification of Normal, Tumor, Hemorrhage and Infarct brain MR images.