Alzheimer's disease detection from structural brain MRI using FFA U-Net segmentation and transfer learning

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that leads to a gradual decline in cognitive and memory function. Multimodal neuroimaging, particularly magnetic resonance imaging (MRI), has become a central diagnostic tool for tracking disease progression, supporting diagnosis, treatment planning, and follow-up monitoring. Manual delineation of brain structures by clinicians remains time-consuming and labor-intensive, and computer-assisted segmentation is complicated by spatial and structural variability as well as intensity inhomogeneity across images. This study proposes an integrated framework for automated brain image segmentation and AD classification to address these challenges.

The framework combines a modified U-Net architecture, termed FFA U-Net, with a fine-tuned VGG16 classifier. FFA U-Net incorporates a residual inception module and a feature fusion attention mechanism into the standard U-Net backbone to perform end-to-end brain tissue segmentation. Segmented outputs are subsequently passed to the fine-tuned VGG16 model for classification of AD status. The framework was evaluated on two datasets, ADNI (843 images) and OASIS (416 images), using subject-wise, non-overlapping training, validation, and test splits to prevent data leakage.

The proposed framework achieved a segmentation Dice Similarity Coefficient (DSC) of 0.929 and a classification accuracy of 98.90%, based on a single data split. These results were consistent across both the ADNI and OASIS datasets, indicating competitive segmentation and classification performance.

The findings suggest that combining attention-augmented segmentation with transfer learning-based classification can effectively support automated AD detection from structural MRI, potentially reducing the manual burden on clinicians. As results are based on a single split, further validation using cross-validation or independent cohorts is warranted to confirm robustness and generalizability. The framework should currently be regarded as an experimental research tool that requires additional clinical validation before deployment in practice.