Reproduce “SkinFlow: Efficient Information Transmission for Open Dermatological Diagnosis via Dynamic Visual Encoding and Staged RL” on OpenResearch

About this paper

Lijun Liu, Linwei Chen, Zhishou Zhang, Meng Tian, Hengfu Cui, Ruiyang Li, Zhaocheng Liu, Qiang Ju and 2 more

A collaborative effort from Baichuan Inc., Peking University First Hospital, and Tsinghua University developed SkinFlow, an AI framework for dermatological diagnosis that optimizes visual information transmission via a dynamic encoder and two-stage reinforcement learning. This system surpassed much larger general-purpose vision-language models in diagnostic accuracy on public and internal datasets while using significantly fewer parameters.

Key insights

• Optimizing geometric capacity and information flow through specialized architectural designs and training strategies is more effective for precision in domain-specific medical AI than merely scaling model parameters.
• The Virtual-Width Dynamic Vision Encoder can effectively overcome the 'capacity collapse' of static vision encoders by implicitly expanding the representational space, allowing the model to adaptively capture subtle visual nuances.
• A two-stage reinforcement learning approach, which first aligns explicit linguistic descriptions and then refines diagnostic predictions, enhances robustness and clinical utility by better distinguishing between describable and non-describable visual features.

Summary by alphaXiv.