Medical Image Segmentation Based on Artificial Intelligence Approaches

Abstract

Medical image segmentation is a critical preprocessing step in computer-aided diagnosis, treatment planning, and biomedical research. While Fuzzy C-Means (FCM) clustering is a widely adopted technique for this task due to its ability to handle inherent ambiguities in medical data, its performance is highly sensitive to initial parameters and is prone to convergence to local optima. This thesis presents a comprehensive approach to overcoming these limitations through two primary contributions. First, we provide an overview of medical imaging and the fundamental challenges of segmentation. We then detail traditional clustering-based methods, with a specific focus on the FCM algorithm, outlining its strengths and well-documented limitations. To address these limitations, we explore bio-inspired optimization metaheuristics as a powerful strategy for guiding the clustering process. The core contribution of this work is the novel hybridization of the Artificial Bee Colony (ABC) algorithm with FCM. The proposed method focuses on the simultaneous optimization of the crucial FCM parameters: primarily the number of cluster centers and their values and the optimization of the objective function by escaping to the local optima, to achieve a superior and more robust segmentation outcome. The effectiveness of this hybrid ABC-FCM approach is rigorously validated through experiments on both simulated brain MRI and real clinical MRI brain images. Results demonstrate a significant improvement in segmentation accuracy and convergence behavior compared to standard FCM and other optimization-enhanced variants. The second major contribution is the development of a new cluster validity index (CVI) to automatically determine the optimal number of segments. This index is designed to enhance the separation metric of the IMI index by incorporating a measure based on Kullback-Leibler (KL) divergence, which better captures the statistical distance between fuzzy clusters. Experimental results confirm that the proposed KL-based CVI outperforms existing indices in accurately identifying the true number of clusters in both synthetic and complex medical imagery. This thesis offers significant advancements in AI-driven medical image segmentation by introducing an optimized clustering framework and a more robust validation metric, both contributing to higher diagnostic reliability.