Swin Transformer-Based Goal-Conditioned Reinforcement Learning for Early Alzheimer's Disease Detection via MRI Classification

Early detection of Alzheimer’s disease (AD) through MRI analysis is critical for timely intervention but remains challenging due to subtle early-stage biomarkers. This paper proposes a novel framework combining a Swin Transformer with goal-conditioned reinforcement learning (GCRL) to enhance diagnostic accuracy. Our GCRL agent learns to classify AD progression stages by iteratively refining predictions across longitudinal MRI sequences, achieving an absolute improvement of 97.93% in accuracy over conventional CNN approaches on the ADNI a dataset. The shifted window mechanism in the Swin Transformer enables efficient multi-scale feature learning, while the goal-conditioned reward structure allows adaptive policy learning for different diagnostic targets. Experimental results demonstrate 97.93% classification accuracy with 99% in precision differentiating early AD from healthy controls, significantly reducing diagnostic variability compared to human expert assessments.