基于深度卷积神经网络的泌尿系结石成分输尿管镜图像诊断模型构建  

作　　者：陈琼秋 孔祥辉 陈合益 方崇国 陈武 陈大可 徐晓敏 

机构地区：[1]浙江省温州市人民医院,325200

出　　处：《浙江临床医学》2025年第2期243-246,共4页Zhejiang Clinical Medical Journal

基　　金：温州市科技计划项目(Y2023451)。

摘　　要：目的采用深度卷积神经网络(CNN)构建用于诊断泌尿系结石成分的输尿管镜(URS)图像分析模型。方法收集2022年1月至2024年7月本院800例接受泌尿系结石URS手术治疗患者的资料,经过筛选,最终获得2475张高质量URS图像数据,随机分为训练集(70%)和测试集(30%)。采用在ImageNet数据集上预训练的Inception v3、ResNet50、AlexNet、VGG 19、DenseNet等网络架构,通过迁移学习技术构建了泌尿系结石成分分析模型。此外,还比较各模型的分类性能,并与泌尿外科医师在术中URS下的评估结果进行对比。结果在训练集和测试集上对构建的泌尿系结石成分URS图像诊断模型进行评估发现,Inception v3、ResNet50、AlexNet、VGG 19、DenseNet模型均具有较高的分类能力。其中Inception v3模型表现最佳,具有最高的准确度(训练集98.10%,测试集98.00%)、AUC值(训练集0.852,测试集0.834)、特异度(训练集82.42%,测试集81.37%)及敏感度(训练集88.36%,测试集86.43%)。一致性检验结果表明,各泌尿系结石成分URS图像诊断模型与医师经验诊断具有较好的一致性,并且Inception v3模型的分类一致性最佳(P<0.001)。结论深度学习技术在泌尿系结石成分诊断中显示出一定的应用潜力。基于CNN构建的泌尿系结石成分URS图像诊断模型具有较好的分类能力,可用于预测泌尿系结石成分。Objective To construct an ureteroscopy(URS)image analysis model for diagnosing the composition of urinary tract stones using deep convolutional neural networks(CNNs).Methods The data were collected from 800 patients with urinary tract stones who underwent URS surgery at our hospital from January 2022 to July 2024.After screening,2475 high-quality URS images with no obvious repetition were obtained.These images were randomly divided into training(70%)and testing(30%)sets.Network architectures were utilized,such as Inception v3,ResNet50,AlexNet,VGG19,and DenseNet,which were pre-trained on the ImageNet dataset,and a model for analyzing the composition of urinary tract stones through transfer learning were constructed.Additionally,the classification performance of each model were compared with the intraoperative URS assessments made by senior urologists.Results Evaluation of the constructed urinary tract stone composition URS image diagnosis model on both the training and testing sets revealed that the Inception v3,ResNet50,AlexNet,VGG19,and DenseNet models all had high classification capabilities.Among them,the Inception v3 model performed the best,with the highest accuracy(98.10%in the training set,98.00%in the testing set),AUC values(0.852 in the training set,0.834 in the testing set),specificity(82.42%in the training set,81.37%in the testing set),and sensitivity(88.36%in the training set,86.43%in the testing set).Consistency test results indicated that the URS image diagnosis models for urinary tract stone composition had good consistency with the diagnostic assessments of experienced urologists,with the Inception v3 model showing the best classification consistency(P<0.001).Conclusion Deep learning technology shows potential in the diagnosis of urinary tract stone composition.The URS image diagnosis model for urinary tract stone composition,built based on CNN,has good classification capabilities and can be used to predict the composition of urinary tract stones.

关 键 词：深度卷积神经网络 泌尿系结石 输尿管镜图像 诊断模型 

分 类 号：R691.4[医药卫生—泌尿科学]