Dynamic Decision Learning Improves Rare Disease Detection

The article presents Dynamic Decision Learning (DDL), a novel approach aimed at improving clinical abnormality grounding for rare diseases. It addresses challenges posed by data scarcity and unstable single-pass inference in supervised fine-tuning methods. By employing frozen large vision-language models (LVLMs), DDL optimizes instructions and consolidates predictions under visual perturbations, thereby enhancing localization quality and generating a consensus-based reliability score to quantify model confidence, demonstrating substantial improvements in brain imaging benchmarks.