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

AI / informaticsAsia-Pacific journal of ophthalmology (Philadelphia, Pa.)· Jun 2026

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

Tsun Hei Yip, Pusheng Xu, Zirong Liu … Xiu Juan Zhang

Read paper

Cataract & RefractiveAI / informaticsQ1IF 4.1

Fine-tuned InternVL3-8B model matched top proprietary vision-language models in myopic maculopathy visual question answering accuracy

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

Tsun Hei Yip … Xiu Juan Zhang

Asia-Pacific journal of ophthalmology (Philadelphia, Pa.) · Jun 2026 · Q1 in Ophthalmology

Purpose

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

Methods

Cross-sectional study.

n = 2591 CFPs and 19,648

Results

fine-tuned model answered about three in four eye-scan questions correctly

0.746accura

Fine-tuned InternVL3-8

0.724accura

Gemini 3 Pro

0.596accura

Claude Sonnet 4.5

0.566accura

Qwen3-VL-30B-A3B-Instr

More results

MM-VQA comprises 2591 colour fundus photographs and 19,648 question-answer pairs.

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).

More results

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).

“

Conclusion

This study provides a valuable visual question answering dataset for myopic maculopathy, supporting the development of disease-specialised vision-language models in ophthalmology.

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

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

Related papers