New framework trims causal graphs to boost medical VQA model generalization

Researchers have developed a new framework called Learnable Causal Trimming (LCT) to improve the generalization of medical Visual Question Answering (MedVQA) models. This approach integrates causal pruning directly into the end-to-end optimization process. LCT utilizes a Dynamic Anatomical Feature Bank (DAFB) to capture common patterns and a differentiable trimming module that suppresses features correlated with these global prototypes, thereby encouraging the model to focus on instance-specific evidence. Experiments on several medical VQA datasets show that LCT enhances robustness and generalization compared to existing debiasing methods. AI

IMPACT Enhances robustness and generalization of medical VQA models by prioritizing causal signals over spurious correlations.

RANK_REASON The cluster contains a research paper detailing a new framework and experimental results. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Dynamic Anatomical Feature Bank (DAFB)
• Learnable Causal Trimming (LCT)
• Medical Visual Question Answering (MedVQA)
• PathVQA
• SLAKE-CP
• VQA-RAD
• Zibo Xu

• paper
• model release
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →