Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

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

Source

arXiv

arxiv.org/abs/2607.15047v1

Read the source

Source

Plain-Language Brief

Researchers propose a parameter-efficient framework for detecting Mild Cognitive Impairment (MCI) using a frozen DINOv2-Small model with adaptive focal loss. The new method utilizes learnable prompt tokens, shared cross-attention layers, and task-conditioned embeddings to achieve spatial explainability directly through attention maps. A modified focal loss is introduced to handle boundary ambiguity, outperforming a computationally heavier ResViT baseline in MCI-class F1.

What the Document Says

  • The research focuses on Mild Cognitive Impairment (MCI), an early stage of cognitive decline that often leads to Alzheimer's disease. The authors note data scarcity, class imbalance, and diagnostic ambiguity as challenges in automating its detection.
  • They propose a parameter-efficient framework using a frozen DINOv2-Small model with three modality-specific learnable prompt tokens and 1.19 million trainable parameters.
  • Each token serves as a query in a shared cross-attention layer over the source image patch tokens, providing spatial explainability through these attention maps.
  • Task-conditioned embeddings are fused via an attention module to quantify modality-level importance per subject.
  • To handle boundary ambiguity, they introduce a MoCA-adapted focal loss that integrates continuous cognitive scores into the training target, loss modulation, and adaptive sample weighting.
  • The proposed architecture outperformed a computationally heavier ResViT baseline in MCI-class F1 by 0.110 under stratified five-fold cross-validation.

Key Concepts Mentioned

  • Mild Cognitive Impairment
  • DINOv2-Small model
  • learnable prompt tokens
  • shared cross-attention layers
  • task-conditioned embeddings
  • focal loss
  • ResViT baseline

Claims Made by the Document

  • Data scarcity, class imbalance, and diagnostic ambiguity are challenges in automating MCI detection.
  • The proposed framework outperforms a computationally heavier ResViT baseline.

Evidence Used

  • The text states the existing challenges in MCI detection
  • The authors compare their proposed architecture with a ResViT baseline and show better performance

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.