基于多模态卷积神经网络的蕈样肉芽肿与五种炎症性皮肤病早期诊断模型
本研究基于临床信息、临床图像和皮肤镜图像构建多模态AI深度学习模型,单中心回顾性纳入1157例蕈样肉芽肿(MF)和炎症性皮肤病病例,含2452张临床图像和6550张皮肤镜图像。结果显示AI模型分类MF与其他炎症性皮肤病的总准确率、精确率、敏感性和特异性均高于参与研究的皮肤科医生,医生+AI组的平均准确率、敏感性和特异性分别达82.9%、86.2%和96.5%,显著高于医生单独组的71.5%、74.6%和94.1%。
要点
- AI模型对MF与炎症性皮肤病的分类性能优于受训皮肤科医生。
- 医生+AI协作可将平均准确率从71.5%提升至82.9%,敏感性从74.6%提升至86.2%。
- 多模态信息(临床资料、临床图、皮肤镜图)可辅助早期MF无创诊断。
核心结论
- 基于多模态信息的AI深度学习模型对早期MF与常见炎症性皮肤病具有较强判别能力,可辅助皮肤科医生提高诊断准确率。
研究设计与样本
- 单中心回顾性研究,基于临床信息、临床图像和皮肤镜图像构建多模态AI模型。
- 共收集1157例MF和炎症性皮肤病病例,包括2452张临床图像、6550张皮肤镜图像及相应临床数据。
- 采用RegNetY400MF作为骨干网络提取皮肤镜和临床图像特征。
- 对比三组:23名接受过10小时以上系统皮肤镜培训的皮肤科医生、AI模型、皮肤科医生+AI模型。
主要发现
- AI模型在MF与其他炎症性皮肤病分类中的总准确率、精确率、敏感性和特异性均高于参与研究的皮肤科医生。
- 医生+AI组平均准确率82.9%、敏感性86.2%、特异性96.5%,医生单独组分别为71.5%、74.6%、94.1%。
- 多分类模型对每种疾病的诊断也更准确。
关键数值与推荐
- 医生+AI组平均准确率82.9%、敏感性86.2%、特异性96.5%。
- 医生单独组平均准确率71.5%、敏感性74.6%、特异性94.1%。
临床怎么用
- AI模型可作为辅助工具帮助皮肤科医生提高早期MF与常见炎症性皮肤病的诊断准确性。
- 临床可结合患者病史、临床图像和皮肤镜图像的多模态信息进行AI辅助诊断。
局限性
- 本研究为单中心回顾性研究。
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
BACKGROUND: Mycosis fungoides (MF) is the most common type of cutaneous T-cell lymphoma, and early-stage MF is difficult to differentiate from erythematous inflammatory disease. With the exception of biopsy, noninvasive information such as a patient's medical history and clinical and dermoscopic images is of great significance for early diagnosis of MF. However, there is a lack of diagnostic models based on convolutional neural networks that can use multimodal information. OBJECTIVES: To develop an artificial intelligence (AI) deep learning model based on multimodal information, to verify its classification efficiency and to construct an AI-aided early diagnostic model of MF and inflammatory skin diseases for dermatologists. METHODS: This was a single-centre retrospective study based on multimodal information, including clinical information, clinical images and dermoscopic images. A total of 1157 cases of MF and inflammatory diseases were collected, including 2452 clinical images, 6550 dermoscopic images and corresponding clinical data. To assess the practicality of using AI models to help with clinical diagnoses, we carried out a comparative study involving three distinct groups: (i) dermatologists, (ii) the AI model and (iii) dermatologists + AI model. The dermatologist group comprised 23 dermatologists with a certain level of expertise and more than 10 h of systematic dermoscopy training. We used RegNetY400MF as the backbone network to extract features from the dermoscopic and clinical images. RESULTS: The AI model demonstrated higher levels of total accuracy, precision, sensitivity and specificity in the classification of MF and other inflammatory skin diseases than participating dermatologists. A significant enhancement was noticed in the average accuracy, sensitivity and specificity for MF and inflammatory diseases in the 'dermatologist + AI' group, with values of 82.9%, 86.2% and 96.5%, respectively, compared with 71.5%, 74.6% and 94.1%, respectively, in the 'dermatologist-only' group. A more accurate diagnosis of each disease was also achieved by the multiclassification model. CONCLUSIONS: The results indicate that our AI model has a significantly strong discriminative ability to assist dermatologists with improving diagnostic accuracy in early-stage MF and common inflammatory skin diseases. Mycosis fungoides (MF) is the most common type of cutaneous lymphoma, which is a rare form of skin cancer. It can be hard to tell early-stage MF apart from other skin problems like eczema, psoriasis and lichen planus. Because of this, doctors need better ways to diagnose MF. Our research team in China created an AI tool to help diagnose early-stage MF and similar skin diseases. The tool used patients’ medical history, skin photos and pictures taken using a magnifying tool (‘dermoscopy’). We studied more than 1,000 patients with MF or inflammatory skin conditions. We reviewed almost 2,500 photos, over 6,500 dermoscopy images, and other medical details. We found the AI model was more accurate at diagnosing MF and these other diseases than the doctors in our study. Our AI model could tell MF apart from other inflammatory skin conditions. This could help doctors quickly diagnose early MF and tell it apart from common skin diseases.
引用
- Liu Z, Zhang Y, Wang K, et al. Early diagnosis model of mycosis fungoides and five inflammatory skin diseases based on a multimodal data-based convolutional neural network. Br J Dermatol. 2025. PMID: 40465821. doi:10.1093/bjd/ljaf212
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。