Hybrid deep learning framework for magnetic resonance imaging-based classification of Alzheimer’s disease

Alzheimer’s disease and related dementias are increasing rapidly, presenting substantial public health challenges due to complex diagnostic protocols and the necessity for prolonged care. Conventional clinical assessments and manual magnetic resonance imaging (MRI) analyses are inherently time-intensive, prone to subjectivity, and susceptible to interobserver variability, constraining timely and accurate diagnosis. To mitigate these limitations, this study introduced a hybrid deep learning framework integrating convolutional neural networks (CNNs) with self-attention mechanisms to improve model interpretability and diagnostic accuracy. For the cross-sectional Open Access Series of Imaging Studies (OASIS)-1 cohort, a 2-dimensional DenseNet-121 backbone followed by a lightweight multi-head Transformer encoder was deployed, enabling efficient slice-level feature extraction and global context modeling. The architecture, tailored to the OASIS-2 longitudinal MRI dataset, employed a 3-dimensional (3D) DenseNet structure augmented with self-attention blocks to enhance volumetric brain feature extraction and capture long-range dependencies. Data augmentation techniques such as random rotation, flipping, and CutMix, alongside regularization techniques including dropout, label smoothing, and early stopping, are utilized to improve generalization and address class imbalance. For the OASIS-1 dataset, the model achieved 91.67 ​% accuracy, with 100 ​% precision, 85.71 ​% sensitivity, and an F1-score of 92.31. The proposed hybrid model achieved superior classification performance on the OASIS-2 dataset, attaining 97.33 ​% accuracy, precision, and sensitivity, along with an F1-score of 98.51 ​%. These results exceeded those of existing baseline models, demonstrating the efficacy of the hybrid approach in accurately delineating cognitive impairment stages based on longitudinal MRI data. The integration of 3D CNNs with attention mechanisms enhances the robustness of Alzheimer’s disease classification, facilitating their application in real-time automated diagnostic tools within clinical neuroimaging workflows. This study presents a robust and explainable framework designed to support clinicians in the early detection and intervention of Alzheimer’s disease.