A trustworthy anatomically-guided region-aware deep learning framework for automated Alzheimer’s disease classification from brain MRI

Alzheimer’s disease (AD) is a neurodegenerative condition that gradually deteriorates the function of cognitive abilities accompanied by irreversible degeneration of brain tissues. Accurate AD diagnoses based on brain MRI are still complicated by the small-scale anatomical changes of the brain, subtocol variability, structural complexity, and overlapped features between the various stages of the disease. In consequence, it makes difficulties in discriminative identification of neurodegenerative biomarkers, reliable classification, weak generalizability, and low confidence for automated diagnostic algorithms. To cope with these problems, this study proposes a novel framework of Trustworthy Anatomically-guided Intelligent Region-aware Alzheimer’s Disease (TAIRA), which is based on region-aware feature representation, adaptive learning, and explainable intelligence in an end-to-end manner. Proposed Attention-guided Region-Aware Residual Dense Network (ARA-RDNet) automatically focuses on sensitive brain regions and extracts discriminative neurodegenerative features with multi-scales by utilizing the methods of spatial attention, channel attention, residual dense connection, and context feature fusion. Moreover, an Adaptive Flatness Guided Convergence Optimizer (AFGCO) is devised for increasing convergence stability, minimizing overfitting problems, and enhancing generalization capabilities. In order to increase the transparency of the model, a Grad-CAM + + explainability approach is utilized, which is validated by conducting perturbation stability testing, faithfulness testing, and localization consistency testing. Classification accuracy on the AA MRI and OASIS datasets was reported as 98.52% and 98.79%, respectively. These results reveal that the proposed approach is capable of recognizing disease-related anatomical patterns, providing reliable Alzheimer’s disease classification.