Multisite T1-weighted MRI classification of Alzheimer’s disease using 3D-CNN-HSCAM architecture with contrastive domain adaptation

Alzheimer’s Disease (AD) presents a significant diagnostic problem due to the considerable diversity in imaging data over several clinical settings. This study presents a new architecture based on Convolutional Neural Networks (CNN) to minimize variability in AD classification. Our model integrates a Hybrid Spatial-Channel Attention Mechanism (HSCAM) with contrastive learning, targeting the challenge of consistent AD diagnosis across four diverse Magnetic Resonance Imaging (MRI) domains. The HSCAM enhances the model’s capability to focus on salient features by adjusting both spatial and channel-wise feature representations, facilitating the extraction of intricate global and local patterns critical for accurate AD detection. Simultaneously, incorporating contrastive learning enables extracting domain-invariant features, significantly boosting the model’s efficacy on unseen datasets. We validated our approach using four classical machine learning classifiers to demonstrate the enhanced feature quality and robustness. Results indicate a marked improvement in classification accuracy, achieving 98.33% accuracy on AD classification, demonstrating a 1.35% improvement over state-of-the-art methods, and a reduction in variability by 1.28% when tested across multiple imaging protocols. This dual-enhancement approach not only sets an innovative mark for AD classification models but also offers substantial potential for application in real-world clinical settings, where imaging protocol variability hinders diagnostic consistency. To ensure clinical relevance, we provided visualizations highlighting influential brain regions in the model’s decisions.