An Improved Deep Learning Unsupervised Approach for MRI Tissue Segmentation for Alzheimer's Disease Detection

Alzheimer's disease (AD) ranks as the sixth leading cause of death, emphasizing the need for early-stage prediction to prevent its progression. Due to the complexity and heterogeneity of medical tests, manually comparing, visualizing, and analyzing data is often difficult and time-consuming. As a result, a computational approach for accurately predicting brain changes through the classification of magnetic resonance imaging (MRI) scans becomes highly valuable, though challenging. This paper introduces a novel method for diagnosing the early stages of AD by utilizing an efficient mapping technique to differentiate between affected and normal MRI scans. The approach combines a hybrid unsupervised learning framework, specifically the adaptive moving self-organizing map (AMSOM) method integrated with Fuzzy K-means. To ensure optimal feature extraction, we introduce a hybrid learning framework that embeds feature vectors in a subspace. The analysis compares various mapping approaches to identify features linked to Alzheimer's disease. The proposed method achieves a classification accuracy of 95.75% on the Open Access Series of Imaging Studies (OASIS) MRI brain image database, outperforming existing methods.