Parameter-efficient Prompt Tuning of Vision Foundation Model With Adaptive Focal Loss for Interpretable MCI Screening

2026-07-16

AI Summary

This study addresses key challenges in automated screening for mild cognitive impairment (MCI)—including data scarcity, class imbalance, and ambiguous clinical boundaries—by proposing a parameter-efficient prompt tuning approach based on a frozen DINOv2-Small backbone. The method introduces three modality-specific learnable prompts that query image patches through a shared cross-attention layer to generate intrinsically interpretable attention maps. To handle boundary ambiguity and enable soft-label generalization, it further incorporates task-conditioned embeddings and a MoCA score–driven adaptive focal loss. This work represents the first integration of prompt tuning with intrinsic interpretability in MCI detection, achieving an MCI-class F1 score of 0.641 (an improvement of 0.110 over ResViT) and an AUC of 0.795 under stratified five-fold cross-validation.

Abstract

Mild Cognitive Impairment is a critical early stage of cognitive decline that frequently precedes Alzheimer's disease, yet its automated detection from neuropsychological drawing tests remains fundamentally constrained by data scarcity, class imbalance, and diagnostic ambiguity near clinical boundaries. Existing methodologies attempt to bypass these constraints using computationally expensive, fully fine-tuned hybrid architectures that relegate spatial explainability to a post-hoc approximation rather than an intrinsic model property. We propose a parameter-efficient framework utilizing frozen DINOv2-Small model adapted via three modality-specific learnable prompt tokens while Operating with 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; as a structural consequence of the architecture. Then task-conditioned embeddings fused via an attention module to quantify modality-level importance per subject. To handle boundary ambiguity, a MoCA-adapted focal loss introduced 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 of 0.641 and an AUC of 0.795, outperforming the computationally heavier ResViT baseline by 0.110 in MCI-class F1.

Problem

• Mild Cognitive Impairment
• data scarcity
• class imbalance
• diagnostic ambiguity
• spatial explainability

Innovation

• parameter-efficient prompt tuning
• adaptive focal loss
• spatial explainability
• vision foundation model
• MCI screening