鲍温病和脂溢性角化人工智能病理诊断模型的建立和评估

Establishment and evaluation of an artificial intelligence model in pathological diagnosis of Bowen’s disease and seborrheic keratosis

目的建立一种皮肤肿瘤人工智能(AI)病理诊断模型并评估其诊断效能。方法选择2种皮肤常见肿瘤鲍温病和脂溢性角化病(SK)作为目标疾病,通过人工标注组织病理H-E切片中的病变区域,为AI提供训练集和验证集。采用AI中基于深度学习的两阶段诊断框架(patch诊断和slide诊断)对此进行综合判断,从而建立相应疾病的诊断模型。选择未经标注病变区域的组织病理H-E切片为AI提供测试集,验证该诊断模型的准确度,运用ROC曲线评价其诊断和鉴别诊断效能。结果第一阶段patch诊断中,Efficientnet_b6模型在patch特征分类上效果更佳,训练集和验证集的灵敏度分别达到94.67%(6680/7056)和95.79%(751/784)。在第二阶段slide诊断中,半结构化数据模型(SSDM)在patch特征聚合方面效果更佳,其训练集特异度为95.00%(6703/7056),验证集特异度为95.28%(747/784);而金融服务数据模型(FSDM)的训练集特异度为91.16%(6432/7056),验证集特异度为82.78%(649/784)。将两阶段诊断模型应用在测试集中,鲍温病和SK的测试准确度分别为92.65%(63/68)和99.21%(126/127)。绘制两阶段诊断模型诊断鲍温病和SK的ROC曲线,AUC值分别为0.97826和0.98698;使用微平均、宏平均2种方式绘制总体ROC曲线,AUC值分别为0.98989和0.98354。结论本研究提出的AI两阶段诊断模型在鲍温病和SK这2种常见皮肤肿瘤的组织病理H-E切片中有较高的诊断及鉴别诊断效能。

Objective To establish an artificial intelligence(AI)model in the pathological diagnosis of skin tumors and to evaluate its diagnostic efficacy.Methods Two common skin tumors,Bowen’s disease and seborrheic keratosis(SK),were selected as the target diseases.Training set and validation set were provided for AI by manually labeling the lesion areas in histopathological hematoxylin-eosin(H-E)stained sections.The two-stage diagnostic framework(patch diagnosis and slide diagnosis)based on deep learning in AI was used to make a comprehensive judgment,so as to establish the diagnostic model of the corresponding diseases.Histopathological H-E sections without labeling lesion areas were selected to provide test set for AI to verify the accuracy of the diagnostic model.Receiver operating characteristic(ROC)curve was used to evaluate the diagnostic and differential diagnostic efficacy.Results In the first stage of patch diagnosis,the Efficientnet_b6 model had a better effect for patch feature classification,and the sensitivity of the training set and validation set reached 94.67%(6680/7056)and 95.79%(751/784),respectively.In the second stage of slide diagnosis,the semi-structured data model(SSDM)was more effective in patch feature aggregation,and the specificity of the training set and the validation set was 95.00%(6703/7056)and 95.28%(747/784),while the specificity of the training set and the validation set of financial service datamodel(FSDM)was 91.16%(6432/7056)and 82.78%(649/784).When the two-stage diagnostic model was applied to the test set,the accuracy of Bowen’s disease and SK was 92.65%(63/68)and 99.21%(126/127),respectively.ROC curves of the two-stage diagnostic model for Bowen’s disease and SK were plotted,with the AUC values being 0.97826 and 0.98698,respectively;and the overall ROC curves were plotted using micro-and macro-average,with the AUC values being 0.98989 and 0.98354,respectively.Conclusion The two-stage AI diagnostic model proposed in this study has a higher diagnostic and differential diagnostic efficacy in the histopathological H-E sections of Bowen’s disease and SK.