Testing Different Models for Brain Tissue Segmentation on ISBR18 Dataset
Segmentation of brain tissue is an essential task in medical image analysis, particularly in neuroimaging and disease diagnosis. This study evaluates and compares three major segmentation approaches in the ISBR18 dataset: atlas-based methods, machine learning techniques, and deep learning architectures. The atlas-based Majority Voting method achieved the highest performance within its category with a dice similarity coefficient of 0.8477, utilizing anatomical templates for segmentation. Among machine learning techniques, K-means clustering demonstrated robust performance with 96% classification accuracy, offering computational efficiency despite limitations in spatial resolution. The deep learning U-Net model trained for binary segmentation achieved 93% accuracy, benefiting from its encoder-decoder architecture for precise boundary detection. While traditional atlas-based approaches provide robust anatomical consistency and machine learning methods offer computational advantages, deep learning models show promise in handling complex segmentation tasks. Future research could integrate these approaches to enhance segmentation performance in the ISBR18 dataset and lead to more accurate and reliable brain tissue segmentation for clinical applications.