EVENT DOSSIER
On September 7, 2026, arXiv released the PubMed-Ophtha dataset. This dataset contains 102,023 panels and their subheadings from 15,842 open-access articles from PubMed Central. The study compared the effects of fixed text templates, medical reports, and domain-specific literature on the fine-tuning of眼底 visual-language models. It was found that domain-specific literature achieved the best average performance in 110 clinical tasks, with a linear detection AUROC of 88.63%, which is superior to the 85.68% of medical reports. The study limited the dataset to image quantity, report volume, or irrelevant articles without reducing performance, indicating that the improvement in performance stemmed from domain density. The research team has released the dataset, the fine-tuned model, and the complete generation process.
arXiv:2605.02720v2 发布 PubMed-Ophtha 数据集以对比不同训练数据源对眼底视觉 - 语言模型的影响。该层级数据集包含来自 PubMed Central 15,842 篇开放获取文章的 102,023 个面板及其子标题，具有高密度领域特征。研究者在相同 CLIP 模型上分别使用固定文本模板、医疗报告及该领域文献进行微调，发现特定领域文献在 110 项临床任务中的平均表现最佳，其线性探测 AUROC 达 88.63%，优于医疗报告的 85.68%。限制数据集至眼底图像数量、医疗报告数据量或与评估无关的文章均未降低性能，表明性能提升源于领域密度。研究组已发布该数据集、微调模型及完整生成流程。