Mild cognitive impairment often shows up before Alzheimer's, and one cheap way to screen for it is in how patients perform simple drawing tests. Automating that reading is hard: little data, lopsided classes, and the tricky cases sitting right on the diagnostic fence. Prior systems threw large fine-tuned models at it and bolted explanations on afterward.

This paper keeps the big DINOv2 vision model frozen and trains just three task-specific prompt tokens, about 1.19 million parameters. Each token queries the image through a shared cross-attention layer, so the attention maps are the explanation, built in rather than reconstructed. A modified focal loss folds continuous MoCA cognitive scores into training to handle the borderline cases. Across five-fold cross-validation it reaches an MCI-class F1 of 0.641 and AUC of 0.795, beating a heavier ResViT baseline by 0.110 in F1.

Those numbers are honest about how hard borderline MCI is, so read the paper for the full setup.