MM-Retinal V2 introduces a high-quality multimodal dataset and innovative pretraining approach for enhanced fundus image analysis across multiple modalities.

• Combines three key ophthalmology imaging modalities: CFP, FFA, and OCT in a single dataset
• Implements a novel fundus anatomy-aware pretraining framework to improve model understanding
• Achieves superior performance while requiring less private data than previous approaches
• Enhances transfer learning capabilities for various downstream ophthalmology tasks

This research represents a significant advancement for medical AI, potentially improving diagnostic capabilities and clinical workflows in ophthalmology by creating more generalizable and effective vision-language models.

MM-Retinal V2: Transfer an Elite Knowledge Spark into Fundus Vision-Language Pretraining