MULTIMODAL MRI ANALYSIS FOR EARLY-STAGE ALZHEIMER’S DETECTION USING ATTENTION MECHANISMS

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disease with an increasingly common prevalence all around the world, necessitating the development of an early, accurate, and minimally invasive diagnosis method for such a devastating condition. However, traditional diagnostic methods such as neuropsychological testing and PET scans may have several shortcomings when used in isolation, namely their high cost, invasive nature, or inability to detect early changes in the brain structure. Therefore, this study presents a new deep learning method that utilizes attention to integrate sMRI, fMRI, and DTI data from early-stage AD patients to achieve better results. The use of attention allows the network to concentrate on specific areas of interest within each input mode and select the most valuable modality-specific features. Several extensive experiments conducted on ADNI datasets suggest that the suggested model can produce an accuracy of 97.63%, precision of 97.47%, recall of 97.77%, F1-Score of 97.60%, specificity of 97.90%, and 0.96 MCC, thus, beating the traditional CNN-based, transfer learning models, and ensemble-based methods. The per-class sensitivity and specificity values are consistently high at an average of 0.961 and 0.959, respectively. It suggests that the suggested framework has a good potential to detect Alzheimer's disease in its early stages, including mild cognitive impairment and cognition-normal states. Further experiments performed by applying ablation analysis suggest that the inclusion of spatial and channel attentions, along with multi-modal fusion, can contribute significantly to performance.