Department of Medical System Engineering, Graduate School of Engineering, Chiba University, Chiba, Japan.
Center for Frontier Medical Engineering, Chiba University, Chiba,
Oncology. 2019;96(1):44-50. doi: 10.1159/000491636. Epub 2018 Aug 21.
This study aimed to use convolutional neural network (CNN), a deep learning software, to assist in cT1b diagnosis.
This retrospective study used 190 colon lesion images from 41 cases of colon endoscopies performed between February 2015 and October 2016. Unenhanced colon endoscopy images (520 × 520 pixels) with white light were used. Images included 14 cTis cases with endoscopic resection and 14 cT1a and 13 cT1b cases with surgical resection. Protruding, flat, and recessed lesions were analyzed. AlexNet and Caffe were used for machine learning. Fine tuning of data to increase image numbers was performed. Oversampling for the training images was conducted to avoid impartiality in image numbers, and learning was carried out. The 3-fold cross-validation method was used. Sensitivity, specificity, accuracy, and area under the curve (AUC) values in the receiver operating characteristic curve were calculated for each group.
The results were the average of obtained values. With CNN learning, cT1b sensitivity, specificity, and accuracy were 67.5, 89.0, and 81.2%, respectively, and AUC was 0.871.
Quantitative diagnosis is possible using an endoscopic diagnostic support system with machine learning, without relying on the skill and experience of endoscopists. Moreover, this system could be used to objectively evaluate endoscopic diagnoses.
本研究旨在使用卷积神经网络（CNN），一种深度学习软件，辅助 cT1b 诊断。
本回顾性研究使用了 190 例结肠镜检查的结肠病变图像，这些图像来自于 2015 年 2 月至 2016 年 10 月期间的 41 例结肠镜检查。使用白光未增强结肠镜检查图像（520×520 像素）。图像包括 14 例经内镜切除的 cTis 病例，以及 14 例手术切除的 cT1a 和 13 例 cT1b 病例。分析了隆起性、平坦性和凹陷性病变。使用 AlexNet 和 Caffe 进行机器学习。对数据进行精细调整以增加图像数量。对训练图像进行过采样，以避免图像数量的偏见，并进行学习。使用 3 倍交叉验证方法。计算每组的敏感性、特异性、准确性和接收器工作特征曲线下的面积（AUC）值。
结果为获得值的平均值。通过 CNN 学习，cT1b 的敏感性、特异性和准确性分别为 67.5%、89.0%和 81.2%，AUC 为 0.871。
使用基于机器学习的内镜诊断支持系统，可以在不依赖内镜医生技能和经验的情况下进行定量诊断。此外，该系统可用于客观评估内镜诊断。
