| 本文已被：浏览 1430次 下载 914次 |
码上扫一扫！ |
| 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 |
| 陈虎艳1,李晓鹏2,李乔1,王朵勤1,徐金华1,吕传峰3,南洋3,陈连军1* |
|
|
(1. 复旦大学附属华山医院皮肤科, 上海 200040;
2. 中南大学数学与统计学院, 长沙 410083;
3. 平安科技智慧医疗影像二队, 上海 200030
*通信作者) |
|
| 摘要: |
| 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%（6 680/7 056）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%（6 703/7 056），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%（6 432/7 056），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0.978 26和0.986 98；使用微平均、宏平均2种方式绘制总体ROC曲线，AUC值分别为0.989 89和0.983 54。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。 |
| 关键词: 人工智能 Bowen病 脂溢性角化病 病理诊断 |
| DOI：10.16781/j.0258-879x.2021.03.0243 |
| 投稿时间：2020-12-09修订日期：2021-02-03 |
| 基金项目:上海市卫生和计划生育委员会智慧医疗专项（2018ZHYL0203）. |
|
| Establishment and evaluation of an artificial intelligence model in pathological diagnosis of Bowen's disease and seborrheic keratosis |
| CHEN Hu-yan1,LI Xiao-peng2,LI Qiao1,WANG Duo-qin1,XU Jin-hua1,Lü Chuan-feng3,NAN Yang3,CHEN Lian-jun1* |
(1. Department of Dermatology, Huashan Hospital, Fudan University, Shanghai 200040, China;
2. School of Mathematics and Statistics, Central South University, Changsha 410083, Hunan, China;
3. Ping An Technology Smart Medical Imaging Team 2, Shanghai 200030, China
*Corresponding author) |
| Abstract: |
| 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% (6 680/7 056) 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% (6 703/7 056) and 95.28% (747/784), while the specificity of the training set and the validation set of financial service data model (FSDM) was 91.16% (6 432/7 056) 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 0.978 26 and 0.986 98, respectively; and the overall ROC curves were plotted using micro- and macro-average, with the AUC values being 0.989 89 and 0.983 54, 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. |
| Key words: artificial intelligence Bowen's disease seborrheic keratosis pathological diagnosis |
