Fundus-R1: Training a Fundus-Reading MLLM on Public Data

Name: Fundus-R1: Training a Fundus-Reading MLLM on Public Data
Uploaded: 2026-10-21T14:18:15.689470+00:00
Duration: 357 s
Description: Fundus-R1 is a reasoning-enhanced multimodal large language model trained only on public fundus images with image-level labels. It overcomes data limitations by using a RAG method to create knowledge-aware reasoning traces, linking visual findings to ophthalmic knowledge. This approach improves performance on fundus reading benchmarks by providing structured diagnostic guidance.

Fundus-R1 is a reasoning-enhanced multimodal large language model trained only on public fundus images with image-level labels. It overcomes data limitations by using a RAG method to create knowledge-aware reasoning traces, linking visual findings to ophthalmic knowledge. This approach improves performance on fundus reading benchmarks by providing structured diagnostic guidance.

Open the full episode