Parameter-efficient Prompt Tuning of Vision Foundation Model With Adaptive Focal Loss for Interpretable MCI Screening
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
Summary
The authors developed a new, efficient computer model to help detect Mild Cognitive Impairment (MCI), an early stage of memory problems that can lead to Alzheimer's. Their method uses a small, fixed visual model combined with special tokens to focus on important parts of brain test drawings, making the results easier to understand. They also introduced a new way to handle uncertain cases by using continuous cognitive scores during training. Their approach performed better in tests than a more complex existing method. This shows promise for improving early diagnosis with fewer computing resources.
Authors
Javad Khoramdel, Farhad Hoseyni, Amirhossein Nikoofard
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.