基于深度学习算法联合Grad-CAM的宫腔镜子宫内膜病变诊断模型研究  

Research on Diagnosis Model of Endometrial Lesions by Hysteroscopy Based on Deep Learning Algorithm Combined with Grad-CAM

作　　者：曹明亮 尹蜜 王庆彬 朱汉峰 李星[1] 张珺[1] 毛林[1] 穆雪峰 曹敏 马于涛 王健[3] 张燕[1] CAO Mingliang;YIN Mi;WANG Qingbin(Department of Obstetrics and Gynecology,Renmin Hospital of Wuhan University,Wuhan Hubei 430060,China;School of Computer Science,Central China Normal University,Wuhan Hubei 430079,China;School of Computer Science,Wuhan University,Wuhan Hubei 430064,China)

机构地区：[1]武汉大学人民医院妇产科,湖北武汉430060 [2]华中师范大学计算机学院,湖北武汉430079 [3]武汉大学计算机学院,湖北武汉430064

出　　处：《实用妇产科杂志》2024年第5期409-413,共5页Journal of Practical Obstetrics and Gynecology

基　　金：武汉大学人民医院2022年度交叉创新人才项目(编号:JCRCZN-2022-009)。

摘　　要：目的:探讨基于深度学习(DL)算法联合可视化技术梯度加权类激活热图(Grad-CAM)开发的宫腔镜子宫内膜病变诊断模型的有效性。方法:选择2021年6月1日至2022年12月31日在武汉大学人民医院妇科行宫腔镜检查的291例患者的303段宫腔镜视频(4781张图像),采用权重采样的方法,将数据集划分为训练集(3703张)和测试集(1078张)。在对训练集用于模型学习与训练后,选择残差神经网络(ResNet18)和高效神经网络(EfficientNet-B0)两种模型架构对测试集分别采用五类和二类分类任务进行模型验证。以病理组织学为金标准,评估其诊断效能,从而选出最优模型,并将Grad-CAM层嵌入最优模型中,输出宫腔镜图像Grad-CAM。结果:①在五类分类任务中,EfficientNet-B0模型的准确度(93.23%)高于ResNet18模型(84.23%);EfficientNet-B0模型在诊断无不典型性子宫内膜增生、子宫内膜息肉、子宫内膜癌、子宫内膜非典型增生、黏膜下肌瘤5种疾病的曲线下面积(AUC)均稍高于ResNet18模型,两者的AUC几乎都在0.980以上。②在准确度的二类分类任务中和对特异度的评估中,两种模型相似,均在93.00%以上,而EfficientNet-B0模型敏感度(91.14%)明显优于ResNet18模型(77.22%)。③EfficientNet-B0模型联合Grad-CAM算法可识别出图像中异常区域,取活检经病理检查证实,模型输出热力图中标记区域约95%为病灶区域。结论:EfficientNet-B0模型联合Grad-CAM研发的宫腔镜诊断模型具有较高的诊断准确度、敏感度和特异度,在诊断子宫内膜病变方面具有应用价值。Objective:To explore the effectiveness of a hysteroscopic endometrial lesion diagnosis model developed based on deep learning(DL)algorithm combined with gradient-weighted class activation mapping(Grad-CAM)visualization technology.Methods:303 hysteroscopy videos(4781 images)of 291 patients who underwent hysteroscopy examination in the Department of Gynecology,Renmin Hospital of Wuhan University from June 1,2021 to December 31,2022 were selected.The dataset was divided into a training set(3703 images)and a test set(1078 images)by weight sampling method.After the training set was used for model learning and training,two model architectures,residual neural network(ResNet18)and efficient neural network(EfficientNet-B0),were selected to verify the model in the test set by five-class and two-class classification tasks,respectively.Taking histopathology as the gold standard,the diagnostic efficacy was evaluated to select the optimal model,and the Grad-CAM layer was embedded in the optimal model to output hysteroscopy images of Grad-CAM.Results:①In the five-class classification tasks,the accuracy of EfficientNet-B0 model(93.23%)was higher than that of ResNet18 model(84.23%);the area under the curve(AUC)of EfficientNet-B0 model in the diagnosis of five diseases,including atypical endometrial hyperplasia,endometrial polyps,endometrial cancer,endometrial atypical hyperplasia,and submucous myoma,was slightly higher than that of ResNet18 model,and the AUC of both models was almost above 0.980.②In the binary classification task of accuracy and the evaluation of specificity,the two models were similar,both above 93.00%,and the sensitivity of EfficientNet-B0 model(91.14%)was significantly better than that of ResNet18 model(77.22%).③EfficientNet-B0 model combined with Grad-CAM algorithm could identify the abnormal areas in the image.After biopsy and pathological examination,it was confirmed that about 95%of the marked areas in the model′s output heatmap were lesion areas.Conclusions:The hysteroscopy diagnostic model develo

关 键 词：宫腔镜 子宫内膜癌 卷积神经网络 梯度加权类激活热图 深度学习 

分 类 号：R711[医药卫生—妇产科学]