BRAIN LESION DELINEATION AND CLASSIFICATION

Abstract

Brain stroke is a life threatening medical emergency which requires immediate medical care. This may be caused due to the blockage or bursting of brain's blood vessel. Based on the cause, it is named as ischemic or hemorrhagic stroke. Stroke may occur due to a variety of reasons, including high blood pressure, head trauma, cardiovascular disease, family history and transient ischemic attack. It is the leading cause of death in India as compared to other western industrialized countries. Therefore, a computer-aided diagnosis (CAD) system can help the physicians for proper and early diagnosis of stroke. The contents of this thesis are categorized into two parts - rst deals with segmentation and second part focuses on classi cation of brain strokes. In the rst part of the thesis, automatic and semi-automatic systems have been developed for the segmentation of stroke lesion from computed tomography (CT) and magnetic resonance (MR) images. In the second part, new feature extraction methods have been proposed to classify brain CT scan images into two or three categories (e.g. hemorrhage, ischemic and normal). In this thesis, Chapter 1 discusses brain strokes, its types, imaging methods like CT and MR, various challenges in their localization and identi cation, and motivation to do further study in order to overcome those challenges. Segmentation of the region of interest from CT and MR images and their classi cation is a challenging task. Therefore, several techniques have been developed for the segmentation and classi cation of medical images in order to ease the diagnosis process. These techniques are discussed in Chapter i Abstract 2. In this thesis, clustering, thresholding and level set methods have been utilized to segment the stroke lesion. In Chapter 3, two methods have been developed for the segmentation of hemorrhagic stroke lesions from CT scan images. The rst method is based on the fuzzy c-means (FCM) clustering and wavelet based thresholding techniques, and second method is based on a newly proposed distance metric for FCM. Chapter 4 is also based on the segmentation of hemorrhagic stroke lesions, where a newly proposed variant of FCM and distance regularized level set evolution (DRLSE) method have been used to identify the region of interest. The new variant of FCM is used to delineate the stroke. To enhance the segmentation results, the DRLSE method has been utilized. Chapter 5 is based on the segmentation of ischemic stroke from MR images. Initially, the MR images are denoised using wavelet based image denoising technique. Then, two di erent variants of segmentation methods thresholding and random forest with the active contour method of ITK-SNAP have been used to segment the ischemic stroke lesion. Chapter 6 and 7 are on the classi cation of brain strokes by extracting useful features from CT scan images. Feature extraction is the most important part of image classi cation. In this thesis, local, global and deep features have been used to extract meaningful features. In Chapter 6, brain stroke CT images are classi ed into two categories using two di erent methods. The rst method is based on the convolutional neural network (CNN) framework. First, all the CT images are preprocessed using quadtree based image fusion method. Thereafter, the proposed convolutional neural network (P CNN) model is trained on the preprocessed image dataset, which classi es them into two categories. The second method focuses on extracting both local and global features. The local binary pattern (LBP), completed LBP (CLBP) and gray level gray co-occurrence matrix (GLCM) features have been used to extract these useful features and then classify images using di erent classi ers. Chapter 7 proposes two local feature descriptors which can classify images into three categories. The rst descriptor is termed as the local neighbourhood pattern (LNP). It is based on the comparison of diagonal neighbours of the center pixel with the mean of whole image intensities. The other neighbours are calculated by comparing with their preceding neighbouring values. Further, the pattern code is calculated for the center pixel. In this way, the codes are computed for all the image pixels. Finally, ii Abstract 1D histogram of obtained image codes is generated as the feature vector. The second method is based on calculating the mean (M) of whole image intensities and double gradients of local neighbourhoods of a center pixel of the image (I) in both x and y directions. Then, we generate an image B by comparing neighbours with M in order to compare double gradient images with this image. Thereafter, histograms of all the images are generated and nally concatenated to form a single feature vector. The proposed method is termed as the local gradient of gradient pattern (LG2P) descriptor. The experimental results obtained by the proposed methods are compared with several previous methods. These results show that they are better than those with an encouraging performance of segmentation and image classi cation. The overall conclusion of the thesis and future scopes are given in the last chapter.