文献速递银屑病与炎症性皮肤病湿疹皮炎与荨麻疹

空间对齐多模态模型提升银屑病与炎症性皮肤病诊断

Enhancing Diagnosis of Psoriasis and Inflammatory Skin Diseases: A Spatially Aligned Multimodal Model Integrating Clinical and Dermoscopic Images.
PubMed 文献 · J Invest Dermatol · 2025-04-10

本研究提出空间对齐多模态对比学习框架,整合皮肤镜与临床照片,在北京协和医院炎症性皮肤病数据集(含8种炎症性皮肤病)上训练。模型在8分类任务准确率0.822,二分类0.911,优于单模态和简单融合方法;在Derm7pt数据集准确率0.807,超越11种多模态方法;辅助20名皮肤科医生将诊断准确率从0.775提升至0.890(P < .05)。

要点

  • 多模态模型在8种炎症性皮肤病分类中准确率达0.822,二分类达0.911。
  • 辅助诊断使20名皮肤科医生准确率从0.775升至0.890(P < .05)。
  • 110例测试中87例(79.1%)诊断准确率提升,对扁平苔藓、痤疮等改善明显。

核心结论

  • 空间对齐多模态模型整合皮肤镜与临床照片,可有效诊断银屑病和其他炎症性皮肤病。

研究设计与样本

  • 研究提出空间对齐多模态对比学习框架,在北京协和医院炎症性皮肤病数据集(含8种炎症性皮肤病)上训练。
  • 在协和数据集进行8分类和二分类任务,并在公开Derm7pt数据集上与11种先进多模态方法比较。
  • 通过20名皮肤科医生辅助诊断实验和热力图验证模型效果。

主要发现

  • 在协和数据集上,模型8分类准确率为0.822,二分类准确率为0.911,优于单模态模型和简单融合方法。
  • 在Derm7pt数据集上,模型准确率达0.807,超越11种先进多模态方法。
  • 辅助诊断使20名皮肤科医生准确率从0.775提升至0.890(P < .05),110例测试中87例(79.1%)准确率提升。
  • 模型对扁平苔藓、痤疮、酒渣鼻和硬斑病等特定疾病诊断改善明显。
  • 热力图验证显示多数病例模型注意力与关键皮损特征一致。

临床怎么用

  • 该模型可作为辅助工具,整合皮肤镜与临床照片,提升皮肤科医生对银屑病等炎症性皮肤病的诊断准确率。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

Psoriasis is a chronic inflammatory disease with significant physical and psychological impacts. To overcome the limitations of single-modality artificial intelligence models in diagnosing inflammatory skin diseases, we propose a multimodal framework-spatial alignment multimodal contrastive learning-integrating dermoscopic and clinical photographs and trained on the developed Peking Union Medical College Hospital - Inflammatory Skin Diseases dataset, which includes 8 inflammatory skin diseases. On the Peking Union Medical College Hospital - Inflammatory Skin Diseases dataset, our model achieved classification accuracies of 0.822 for an 8-class classification task and 0.911 for a binary classification task, outperforming single-modality models and simple fusion approaches. Furthermore, on the publicly available Derm7pt dataset, the proposed model surpassed 11 state-of-the-art multimodal methods, achieving an accuracy of 0.807. When used for diagnostic assistance, the model significantly improved the diagnostic accuracy of 20 dermatologists, increasing from 0.775 to 0.890 (P < .05). Among 110 test cases, 87 (79.1%) showed improved diagnostic accuracy. The model demonstrated notable enhancements in diagnosing specific conditions, such as lichen planus, acne, rosacea, and morphea. Validation using heatmaps confirmed that the model's attention aligned with key lesion features in most cases. This study introduces a multimodal model that bridges scale discrepancies between dermoscopic and clinical photographs, offering an efficient tool for diagnosing psoriasis and other inflammatory skin diseases.

来源:J Invest Dermatol,查看原文。版权归原作者所有。

引用

  1. Wang J, Zhang Y, Xie F, et al. Enhancing Diagnosis of Psoriasis and Inflammatory Skin Diseases: A Spatially Aligned Multimodal Model Integrating Clinical and Dermoscopic Images. J Invest Dermatol. 2025. PMID: 40216156. doi:10.1016/j.jid.2025.03.034

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。