摘要
目的为克服深度学习模型黑盒不可解释的缺点,本研究旨在探讨可解释性计算机视觉模型在内镜下内痔诊断及危险分级中的应用。方法收集苏州大学附属第一医院内镜中心的肛齿状线上倒镜图片,分为内痔组和正常组;根据LDRf分级标准,对内痔组进一步分级为Rf0、Rf1及Rf2三组。针对有无内痔、红色征、糜烂、血栓及活动性出血,构建基于ResNet50V2的可解释化模型,并利用江苏大学附属金坛医院内镜中心的内镜图片进行外部验证。使用准确性、敏感性、特异性以及F1值等指标对比可解释化模型与传统深度学习黑盒模型的表现,并与两位不同年资内镜医生进行比较。结果ResNet可解释化模型的准确性为0.957、敏感性为0.978、特异性为0.974,F1值为0.958,其准确性高于黑盒模型的0.938,高年资内镜医生的0.933及低年资医生的0.907。此外,模型采用Grad-CAM方法突出图像中对模型推理依据的区域。结论本研究通过收集内镜下肛齿状线上倒镜图像,构建可解释化计算机视觉模型并进行外部验证,提示该模型在内镜下内痔诊断与评级中表现优于传统深度学习黑盒模型。该模型在未来临床内镜诊疗中具有良好应用前景。
Objective To the investigate the application of explainable computer vision models in endoscopic evaluation of internal hemorrhoids to overcome the defect in the uninterpretability of deep learning black-box model.Methods Upper anus dentate line images of the endoscopy center of the First Affiliated Hospital of Soochow University were collected and divided into internal hemorrhoids group and normal group.According to the LDRf grading standard,the internal hemorrhoids group was classified into Rf0,Rf1 and Rf2.For the presence of internal hemorrhoids,red sign,erosion,thrombus and active bleeding,an explainable model based on ResNet50V2 was constructed and externally validated using endoscopic pictures from the Jintan Hospital affiliated to Jiangsu University.Accuracy,sensitivity,specificity,and F1-score were used to evaluate interpretable and black-box models,as well as two endoscopists with different years of experience.Results The accuracy,sensitivity and specificity of ResNet interpretable model were 0.957,0.978,0.974,and F1-score was 0.958,which was higher than 0.938 for black box model,0.933 for senior endoscopists and 0.907 for junior endoscopists.The Grad-CAM promoted model s visualization and explanation.Conclusion The study developed and externally validated an interpretable computer vision model,which showed that the model performed better than the traditional deep learning black-box model in endoscopic diagnosis and grading of internal hemorrhoids.Interpretable deep learning shows promise in the future application in clinical endoscopy.
作者
刘璐
林嘉希
朱世祺
高静雯
刘晓琳
许春芳
朱锦舟
LIU Lu;LIN Jia-xi;ZHU Shi-qi;GAO Jing-wen;LIU Xiao-lin;XU Chun-fang;ZHU Jin-zhou(Department of Gastroenterology,the First Affiliated Hospital of Soochow University,Suzhou,Jiangsu 215000,China)
出处
《现代消化及介入诊疗》
2023年第8期972-975,980,共5页
Modern Interventional Diagnosis and Treatment in Gastroenterology
基金
国家自然科学基金(82000540)
苏州市科技计划(SKY2021038)
苏州市科教兴卫项目(KJXW2019001)。
关键词
可解释性
深度学习
内痔
消化内镜
LDRf分级
梯度加权分类激活映射
interpretability
deep learning
internal hemorrhoids
endoscopy
LDRf standard
Gradient-weighted Class Activation Mapping(Grad-CAM)
