基于深度学习的超声内镜分站和胰腺分割识别系统

A station recognition and pancreatic segmentation system in endoscopic ultrasonography based on deep learning

摘要:

目的 尝试构建1个基于深度学习的内镜超声检查（endoscopic ultrasonography，EUS）质量控制系统，并验证其价值。 方法 从武汉大学人民医院消化内镜中心数据库中，回顾性收集2016年12月—2019年12月间的269个EUS检查资料，分为：（1）用于训练模型的训练数据集A，包含205个检查，其中16 305张图像用于分类训练，1 953张图像用于分割训练；（2）用于评估模型性能的测试数据集B，包含44个检查，其中1 606张图像用于分类验证，480张图像用于分割验证；（3）用于内镜医师与模型进行比较的数据集C，包含20个检查，共150张图像。EUS专家（具有10年以上的EUS操作经验）甲和乙通过讨论对训练集A和测试集B、C的所有图像进行分类和标注，其结果用作金标准。EUS专家丙和高年资EUS医师（具有5年以上的EUS操作经验）丁、戊对数据集C中的图像进行分类和标注，其结果用于与深度学习模型进行比较。主要观察指标包括分类的准确率、分割的Dice（F1指数）和一致性分析的Kappa系数。 结果 在测试数据集B中，模型分类的平均准确率为94.1%，胰腺分割的平均Dice为0.826，血管分割的平均Dice为0.841。在数据集C中，模型的分类准确率达到90.0％，专家丙、高年资医师丁和戊分别为89.3％、88.7％和87.3％；模型的胰腺和血管分割Dice系数分别为0.740和0.859，专家丙分别为0.708和0.778，高年资医师丁分别为0.747和0.875，高年资医师戊分别为0.774和0.789，模型与专家的水平相当。一致性分析结果显示，模型与内镜医师之间达成了较好的一致性（Kappa系数分别为：模型与专家丙间0.823、模型与高年资医师丁间0.840、模型与高年资医师戊间0.799）。 结论 基于深度学习的EUS分站和胰腺分割识别系统可以用于胰腺EUS的质量控制，具有与EUS专家相当的分类和分割识别水平。

Abstract:

ObjectiveTo develop an endoscopic ultrasonography (EUS) station recognition and pancreatic segmentation system based on deep learning and to validate its efficacy. MethodsData of 269 EUS procedures were retrospectively collected from Renmin Hospital of Wuhan University between December 2016 and December 2019, and were divided into 3 datasets: (1)Dataset A of 205 procedures for model training containing 16 305 images for classification training and 1 953 images for segmentation training; (2)Dataset B of 44 procedures for model testing containing 1 606 images for classification testing and 480 images for segmentation testing; (3) Dataset C of 20 procedures with 150 images for comparing the performance between models and endoscopists. EUS experts (with more than 10 years of experience) A and B classified and labeled all images of dataset A, B and C through discussion, and the results were used as the gold standard. EUS expert C and senior EUS endoscopists (with more than 5 years of experience) D and E classified and labeled the images in dataset C, and the results were used for comparison with model. The main outcomes included accuracy of classification, Dice (F1 score) of segmentation and Cohen Kappa coefficient of consistency analysis. ResultsIn test dataset B, the model achieved a mean accuracy of 94.1% in classification. The mean Dice of pancreatic and vascular segmentation were 0.826 and 0.841 respectively. In dataset C, the classification accuracy of the model reached 90.0%. The classification accuracy of expert C, senior endoscopist D and E were 89.3%, 88.7% and 87.3%, respectively. The Dice of pancreatic and vascular segmentation in the model were 0.740 and 0.859, 0.708 and 0.778 for expert C, 0.747 and 0.875 for senior endoscopist D, and 0.774 and 0.789 for senior endoscopist E. The model was comparable to the expert level.Consistency analysis showed that there was high consistency between the model and endoscopists (the Kappa coefficient was 0.823 between model and expert C, 0.840 between model and senior endoscopist D, and 0.799 between model and senior endoscopist E). ConclusionEUS station classification and pancreatic segmentation system based on deep learning can be used for quality control of pancreatic EUS, with a comparable performance of classification and segmentation to that of EUS experts.