Department of Gastroenterology and Hepatology, Osaka University Graduate School of Medicine, Osaka, Japan.
J Gastroenterol Hepatol. 2021 Aug;36(8):2239-2246. doi: 10.1111/jgh.15479. Epub 2021 Mar 10.
The morphological diagnosis of microvessels on the surface of superficial esophageal squamous cell carcinomas using magnifying endoscopy with narrow-band imaging is widely used in clinical practice. Nevertheless, inconsistency, even among experts, remains a problem. We constructed a convolutional neural network-based computer-aided diagnosis system to classify the microvessels of superficial esophageal squamous cell carcinomas and evaluated its diagnostic performance.
In this retrospective study, a cropped magnifying endoscopy with narrow-band images from superficial esophageal squamous cell carcinoma lesions was used as the dataset. All images were assessed by three experts, and classified into three classes, Type B1, B2, and B3, based on the Japan Esophagus Society classification. The dataset was divided into training and validation datasets. A convolutional neural network model (ResNeXt-101) was trained and tuned with the training dataset. To evaluate diagnostic accuracy, the validation dataset was assessed by the computer-aided diagnosis system and eight endoscopists.
In total, 1777 and 747 cropped images (total, 393 lesions) were included in the training and validation datasets, respectively. The diagnosis system took 20.3 s to evaluate the 747 images in the validation dataset. The microvessel classification accuracy of the computer-aided diagnosis system was 84.2%, which was higher than the average of the eight endoscopists (77.8%, P < 0.001). The area under the receiver operating characteristic curves for diagnosing Type B1, B2, and B3 vessels were 0.969, 0.948, and 0.973, respectively.
The computer-aided diagnosis system showed remarkable performance in the classification of microvessels on superficial esophageal squamous cell carcinomas.
使用窄带成像放大内镜对食管浅表鳞状细胞癌表面微血管形态进行诊断在临床实践中得到广泛应用。然而，即使是专家之间也存在不一致的情况。我们构建了一个基于卷积神经网络的计算机辅助诊断系统来对食管浅表鳞状细胞癌的微血管进行分类，并评估其诊断性能。
在这项回顾性研究中，使用来自食管浅表鳞状细胞癌病变的放大窄带内镜图像作为数据集。所有图像均由三位专家进行评估，并根据日本食管学会分类标准分为 B1、B2 和 B3 三种类型。数据集分为训练集和验证集。使用训练集对卷积神经网络模型（ResNeXt-101）进行训练和调整。为了评估诊断准确性，使用计算机辅助诊断系统和八名内镜医生对验证集进行评估。
共有 1777 个和 747 个裁剪图像（共 393 个病灶）分别纳入训练集和验证集中。诊断系统评估验证集中的 747 张图像需要 20.3 秒。计算机辅助诊断系统的微血管分类准确率为 84.2%，高于八位内镜医生的平均准确率（77.8%，P<0.001）。用于诊断 B1、B2 和 B3 型血管的受试者工作特征曲线下面积分别为 0.969、0.948 和 0.973。
计算机辅助诊断系统在食管浅表鳞状细胞癌微血管分类方面表现出优异的性能。
