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

To build a visual question answering (VQA) dataset for fine-tuning and evaluating vision-language models (VLMs) in myopic maculopathy (MM).

Cross-sectional study.

Colour fundus photographs (CFPs) from two publicly available datasets were graded using META-PM classification system. GPT-5 was used to generate clinical captions, true/false [TFQ] and open-ended [OEQ] question-answer pairs, all of which were manually verified. InternVL3-8B was fine-tuned on this dataset and evaluated against Gemini 3 Pro, Claude Sonnet 4.5, Qwen3-VL-30B-A3B-Instruct, and pre-trained InternVL3-8B. OEQ responses was evaluated by GPT-5 using a three-level scoring system (0, completely incorrect; 0.5, partially correct; 1, fully correct) and summarized as weighted accuracy. Overall accuracy was defined as the arithmetic mean of the TFQ and OEQ accuracies.

MM-VQA comprises 2591 CFPs and 19,648 question-answer pairs. Fine-tuned InternVL3-8B model achieved an overall accuracy of 0.746, surpassing Claude Sonnet 4.5 (0.596), Qwen3-VL-30B-A3B-Instruct (0.566), and pre-trained InternVL3-8B (0.428) (all P < 0.001), while showing no significant difference compared with Gemini 3 Pro (0.724, P = 0.642). For TFQ, the fine-tuned model reached an accuracy of 0.919, outperforming Gemini 3 Pro (0.881), Qwen3-VL-30B-A3B-Instruct (0.834), Claude Sonnet 4.5 (0.796), and the pre-trained model (0.696) (all P < 0.001). On OEQ, it also ranked highest (0.572), outperforming Gemini 3 Pro (0.567, P = 0.044), Claude Sonnet 4.5 (0.395, P < 0.001), Qwen3-VL-30B-A3B-Instruct (0.297, P < 0.001) and the pre-trained model (0.160, P < 0.001).

This study provides a valuable VQA dataset for MM, supporting the development of disease-specialised VLMs in ophthalmology.