摘要
目的:构建基于多实例学习的消化道病理辅助筛查模型,帮助病理医生在消化道远程会诊场景下快速准确地进行诊断。方法:从郑州大学第一附属医院国家远程医疗中心的远程病理诊断平台收集从2018年至2023年的病理切片图像,经关键词筛选得到6228张消化道切片图像,经数据清洗最终得到5658张切片图像用于模型的训练与测试。该模型由多个基于深度学习的算法模块构成,包括前景提取、图像块分类和切片分类。模型使用消化道阳性切片679张和阴性切片1304张进行训练,在2353张阳性切片和295张阴性切片上进行测试。该模型输入为消化道数字病理切片,输出为该数字切片为阳性切片的概率值。选取具有基层医院医生诊断的切片1027张作为独立数据集比较本文模型和基层医生的分类性能。结果:本模型切片分类的特异度、敏感度和ROC曲线下面积(AUC)分别为0.902、0.955和0.978。选取1027张切片进行人机结果对比,结果显示,本模型的敏感度为0.96,特异度0.95;与基层医院医生诊断结果相比,敏感度提升6.7%,特异度提升26.7%。结论:基于多实例学习的消化道病理辅助筛查模型准确性较高,可提升基层医疗机构医生诊断水平,为临床诊疗提供有效的决策支持。
Objective To construct a digestive tract pathology assistant screening model based on multi-case learning to facilitate pathologists making rapid and precise diagnoses in digestive tract remote consultation context.Methods Pathological section images from 2018 to 2023 were collected from the remote pathological diagnosis platform of the National Telemedicine Center of the First Affiliated Hospital of Zhengzhou University,6,228 digestive tract section images were obtained after keyword filtering,and 5,658 section images were finally obtained by data cleaning for model training and testing.The model consists of several algorithm modules based on deep learning,including foreground extraction,image block classification and section classification.The model was trained by using 679 positive and 1,304 negative sections of the digestive tract,and was subsequently tested on 2,353 positive and 295 negative sections.The input of the model was digital pathological sections of the digestive tract,and the output was the probability value of the digital sections being positive sections.1,027 sections diagnosed by doctors from primary hospitals were selected as an independent dataset to compare the classification performance of the model with that of doctors from primary hospitals.Results The specificity,sensitivity and AUC(area under ROC curve)of the model in this study were 0.902,0.955 and 0.978,respectively.When subjected to a comparative analysis of human-machine diagnosis utilizing 1,027 slices,the model demonstrated a sensitivity and specificity of 0.96 and 0.95,respectively.Compared with the results of doctors from primary hospitals,the sensitivity increased by 6.7%and the specificity increased by 26.7%.Conclusion The digestive tract pathology assistant screening model based on multi-case learning has high accuracy,which can assist doctors in primary medical institutions to improve their diagnosis and provide effective decision-making support for clinical diagnosis and treatment.
作者
叶明
石金铭
崔芳芳
何贤英
YE Ming;SHI Jinming;CUI Fangfang;HE Xianying(National Engineering Laboratory for Internet Medical Systems and Applications,the First Affiliated Hospital of Zhengzhou University,Zhengzhou 450052,Henan Province,China;Shanghai Artificial Intelligence Laboratory)
出处
《中国数字医学》
2024年第9期77-83,共7页
China Digital Medicine
基金
科技创新2030-“新一代人工智能”重大专项(2022ZD0160704)
河南省重点研发与推广专项(232102311057)
2023年度河南省高等学校重点科研项目(23B330003)。
关键词
多实例学习
消化道病理
远程会诊
深度神经网络
智能辅助筛查
Multi-case learning
Digestive pathology
Remote consultation
Deep neural network
Intelligent assistant screening
