基于U-Net深度学习慢性萎缩性胃炎模型的应用与研究

Application and research of diagnosis model of chronic atrophic gastritis based on U-Net deep learning

目的评估基于深度学习技术中的图像分割模型U-Net构建的慢性萎缩性胃炎诊断模型的诊断性评价指标及其与病理诊断的一致性。方法选取2019年8月1日至2020年8月1日于首都医科大学宣武医院行胃镜检查的慢性萎缩性胃炎患者1711例的胃镜检查胃部图片,利用计算机产生的随机数字方法,选取高质量图片5290张进入研究。根据萎缩严重程度应用分层随机法将70%的图片(3703张)纳入训练集,30%的图片(1587张)纳入测试集。模型方面,采用U-Net网络结构作为基线模型和内部参数初始权重,通过训练集对模型进行重新训练以调整权重,通过测试集检验模型重新训练后的灵敏度、特异度、正确率等指标。结果模型对慢性萎缩性胃炎诊断的灵敏度、特异度、正确率分别为92.73%、92.24%、92.63%,AUC为0.932(95%CI:0.916~0.948)(P<0.001),其与病理诊断的一致性Kappa值为0.796(P<0.001)。测试识别图片中假阳性26张(1.6%),假阴性91张(5.7%)。结论本研究建立了基于U-Net深度学习的慢性萎缩性胃炎诊断模型,并通过回顾性研究发现该诊断模型对慢性萎缩性胃炎的诊断性评价指标良好,并且与病理诊断有较高的一致性。

Objective To evaluate the diagnostic evaluation indexes of the U-Net deep learning-based diagnostic model of chronic atrophic gastritis and its consistency with pathological diagnosis.Methods In this study,gastroscopic pictures of 1711 patients with chronic atrophic gastritis who underwent gastroscopy in Xuanwu Hospital of Capital Medical University from Aug.1st,2019 to Aug.1st,2020 were selected,and 5290 high-quality pictures were selected into the study by using computer-generated random number method.According to the severity of atrophy by stratified random method,70% of the images(3703)were included in the training set and 30% of the images(1587)were included in the test set.In terms of the model,the U-Net network structure was used as the baseline model and the initial weight of internal parameters.The model was retrained through the training set to adjust the weight.The sensitivity,specificity,accuracy and other indexes of the model after retraining were tested through the test set.Results The sensitivity,specificity and accuracy of the model for the diagnosis of chronic atrophic gastritis were 92.73%,92.24% and 92.63%,respectively.The AUC was 0.932(95%CI:0.916-0.948)(P<0.001),and the Kappa value of consistency with pathological diagnosis was 0.796(P<0.001).Among the identified images,26(1.6%)were false positive and 91(5.7%)were false negative.Conclusion In this study,a deep learning-based diagnostic model of chronic atrophic gastritis was established,and through retrospective study,it was found that the diagnostic evaluation index of this diagnostic model for chronic atrophic gastritis was good,and it was consistent with the pathological diagnosis.