Approche robuste pour la segmentation et la classification d’images m´edicale

Abstract

Image segmentation is a vital process in various fields, including robotics, object recognition, and medical imaging. In medical imaging, accurate segmentation of brain tissues from MRI images is crucial for diagnosing and treating brain disorders such as Alzheimer’s disease, epilepsy, schizophrenia, multiple sclerosis, and cancer. This thesis proposes an automatic fuzzy method for brain MRI segmentation. Firstly, the proposed method aims to improve the efficiency of the Fuzzy C-Means (FCM) algorithm by reducing the need for manual intervention in cluster initialization and determining the number of clusters. For this purpose, we introduce an adaptive splitmerge technique that effectively divides the image into several homogeneous regions using a multi-threshold method based on entropy information. During the merge process, a new distance metric is introduced to combine the regions that are both highly similar within the merged region and effectively separated from others. The cluster centers and numbers obtained from the adaptive split-merge step serve as the initial parameters for the FCM algorithm. The obtained fuzzy partitions are evaluated using a novel proposed validity index. Secondly, we present a novel method to address the challenge of noisy pixels in the FCM algorithm by incorporating spatial information. Specifically, we assign a crucial role to the central pixel in the clustering process, provided it is not corrupted with noise. However, if it is corrupted with noise, its influence is reduced. Furthermore, we propose a novel quantitative metric for replacing the central pixel with one of its neighbors if it can improve the segmentation result in terms of compactness and separation. To evaluate the effectiveness of the proposed method, a thorough comparison with existing clustering techniques is conducted, considering cluster validity functions, segmentation accuracy, and tissue segmentation accuracy. The evaluation comprises comprehensive qualitative and quantitative assessments, providing strong evidence of the superior performance of the proposed approach.