ResearchPod Summary
Automated detection of Mild Cognitive Impairment (MCI) from neuropsychological drawing tests (such as clock drawing or cube copying) is often hindered by data scarcity, class imbalance, and the inherent ambiguity of clinical diagnostic boundaries. Existing deep learning approaches often rely on computationally expensive, fully fine-tuned architectures that treat interpretability as a secondary, post-hoc approximation. This paper asks whether a parameter-efficient, foundation-model-based approach can achieve high diagnostic accuracy while providing intrinsic, structural interpretability.
The authors propose a framework that keeps a DINOv2-Small vision foundation model frozen, utilizing only 1.19 million trainable parameters. The architecture processes three neuropsychological drawing tasks (Clock, Cube, and Trail Making) using modality-specific learnable prompt tokens. These tokens act as queries in a shared cross-attention layer, allowing the model to focus on relevant spatial features within each image. To address diagnostic ambiguity near the clinical cutoff (a MoCA score of 25), the authors introduce a MoCA-adapted focal loss that incorporates continuous cognitive scores into the training target, loss modulation, and adaptive sample weighting. The model is evaluated using stratified five-fold cross-validation on a dataset of 918 subjects.
The proposed architecture achieves an MCI-class F1 score of 0.641 and an AUC of 0.795, outperforming the ResViT baseline by 0.110 in F1 score. By using cross-attention weights as a direct byproduct of the forward pass, the model provides spatial saliency maps that highlight which parts of a drawing contributed to the classification. Additionally, the learnable query attention module provides a second layer of interpretability by quantifying the relative importance of each drawing modality (Clock, Cube, or Trail) for a specific patient's risk profile.
This work demonstrates that massive, fully fine-tuned models are not strictly necessary for medical image analysis tasks with limited data. By leveraging frozen foundation models and task-specific prompt tuning, the authors provide a pathway toward building diagnostic tools that are both computationally efficient and inherently interpretable, which is a critical requirement for clinical adoption and trust in AI-assisted medical decision-making.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.