Abstract
Mild Cognitive Impairment (MCI) is a critical early stage of cognitive decline that frequently precedes Alzheimer's disease. However, automated detection from neuropsychological drawing tests remains fundamentally constrained by data scarcity, class imbalance, and diagnostic ambiguity near clinical boundaries. Existing methodologies often bypass these limitations using computationally expensive fully fine-tuned hybrid architectures, relegating spatial explainability to a post-hoc approximation rather than an intrinsic model property.
We propose a parameter-efficient framework utilizing a frozen DINOv2-Small model adapted via three modality-specific learnable prompt tokens, operating with only 1.19 million trainable parameters. Each token serves as a query in a shared cross-attention layer over the source image patch tokens. Crucially, spatial explainability is achieved directly through these attention maps, resulting from the architecture's structural design. Task-conditioned embeddings are then fused via an attention module to quantify modality-level importance per subject.
To address boundary ambiguity, we introduce a MoCA-adapted focal loss that integrates continuous cognitive scores into the training target, loss modulation, and adaptive sample weighting, strictly generalizing standard soft-label approaches. Under stratified five-fold cross-validation, the proposed architecture yields an MCI-class F1 score of 0.641 and an AUC of 0.795, outperforming the computationally heavier ResViT baseline by 0.110 in MCI-class F1.
Blogger's Review: This study not only demonstrates an efficient parameter tuning approach for MCI screening under resource constraints but also enhances model interpretability through the introduction of adaptive focal loss, providing significant insights for future clinical applications.