Benchmarking and fine-tuning vision-language models on a visual question answering dataset for myopic maculopathy

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Myopic maculopathy (MM) poses a significant challenge in ophthalmology.
• Developing advanced tools for diagnosing and monitoring MM is crucial.
• Vision-language models (VLMs) show promise in interpreting medical images.

Purpose of the Study:

• To establish a comprehensive visual question answering (VQA) dataset specifically for myopic maculopathy (MM).
• To facilitate the fine-tuning and evaluation of vision-language models (VLMs) for MM-related tasks.
• To advance AI-driven diagnostic capabilities in ophthalmology.

Main Methods:

• A cross-sectional study was conducted using colour fundus photographs (CFPs).
• A novel MM-VQA dataset was constructed, including clinical captions and question-answer pairs (TFQ and OEQ), generated using GPT-5 and manually verified.
• The InternVL3-8B model was fine-tuned on this dataset and benchmarked against other leading VLMs.

Main Results:

• The MM-VQA dataset contains 2,591 CFPs and 19,648 question-answer pairs.
• The fine-tuned InternVL3-8B achieved a high overall accuracy of 0.746, outperforming several other models.
• The model demonstrated strong performance in both true/false questions (0.919 accuracy) and open-ended questions (0.572 accuracy).

Conclusions:

• The developed MM-VQA dataset is a valuable resource for advancing VLM research in ophthalmology.
• This dataset supports the creation of specialized VLMs for improved myopic maculopathy assessment.
• The findings underscore the potential of fine-tuned VLMs in clinical ophthalmology applications.