A BRAIN-CONNECTED NETWORK MODEL - ACQUIRING MRI STRUCTURAL AND FUNCTIONAL CHARACTERISTICS TO ENHANCE ALZHEIMER’S DISEASE PREDICTION

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that steadily deteriorates brain structure and function, ultimately leading to a significant decline in cognitive abilities. Timely diagnosis may stop disease development and improve patient outcomes, making early detection of AD critical.  For a correct diagnosis of AD, Magnetic Resonance Imaging (MRI) is crucial.  On the other hand, imaging methods necessitate better training and more time.  Automated and precise AD detection has recently been made possible with the help of Deep Learning (DL) methods.  These methods are able to learn intricate patterns from massive amounts of imaging data, picking up on small variations that people would detect. Yet, many existing DL methods focus primarily on single-modality data or shape-based features, often neglecting other complementary information. As a result, they fail to fully capture the complex interactions within the brain, leading to reduced classification performance. This research suggests a new multimodal framework, Enhanced Ensemble Brain Connected Alzheimer’s Disease Network (EEBCADnet), which integrates physical along with functional MRI data, to circumvent this shortcoming.  To create 360 ROIs, the suggested approach mostly uses fMRI data that has been preprocessed along with parcellated utilizing the Glasser atlas. Then, Functional connectivity (FC) matrices are constructed by computing correlations among ROI time series, followed by graph-theoretical feature extraction to capture both local and global network measures. To combine these functional features with structural features, a concatenation strategy is employed which forms a fused feature vector. A fully connected (FC) layer as well as a softmax layer are used to make the final stage-wise AD forecast using these vectors. Experimental results on the ADNI and OASIS-3 datasets show that EEBCADnet consistently outperforms traditional models—including CNN, SVM-ANN, 2DCNN-SVM, CSEPC, and MAGNet, achieving accuracies of 93.58% and 94.23%, respectively. The suggested method is robust and reliable, and paired t-test analysis proves that these enhancements are statistically significant.