认识贝芙汀

这项多中心前瞻性研究把皮肤照片与病历文本一起用于模型训练和验证。临床验证纳入15家医院的6,043例病例，覆盖31种常见皮肤病，重点说明图像与病史信息结合后如何辅助皮肤病识别。研究结果不能代替医生结合病史与面诊作出的诊断。

This prospective multicenter study evaluated a dual-channel model that combines skin images with extracted medical text. The clinical validation set included 6,043 cases from 15 hospitals and covered 31 common skin diseases. DOI: 10.3389/frai.2023.1213620.

研究使用对比学习、聚类和数据过滤处理在线皮肤病社区中的噪声标签，原始数据包含超过20万张图像和22万条医疗记录，并围绕22种常见皮肤病进行建模。它展示了互联网数据在皮肤图像研究中的潜力，但模型表现仍受数据来源、标注质量和验证场景限制。

The study explored how contrastive learning, clustering, and data filtering can reduce noisy labels in online dermatology images. The source dataset contained more than 200,000 images and 220,000 medical records; the reported results should be interpreted within the study datasets and validation settings. DOI: 10.1038/s41746-024-01014-x.

研究把图像质量控制、皮损分类、定位、计数与分割放入同一套痤疮识别流程。研究数据中的皮损分类准确率为83%，并报告了不同检测与分层任务的表现。数字用于理解该研究的测试结果，不能直接换算为个人测肤或临床诊断准确率。

This study placed image-quality control, lesion classification, localization, counting, and segmentation into one acne-detection workflow. In the study dataset, lesion classification accuracy was 83%, while performance varied by task and validation setting. DOI: 10.3389/fmed.2024.1344314.