使用非放大内窥镜白光图像进行早期结肠癌的计算机辅助诊断。
Computer-Aided Diagnosis of Early-Stage Colorectal Cancer Using Non-Magnified Endoscopic White Light Images.
发表日期：2023 Feb 02
作者：
Daiki Nemoto, Zhe Guo, Shinichi Katsuki, Takahito Takezawa, Ryo Maemoto, Keisuke Kawasaki, Ken Inoue, Takashi Akutagawa, Hirohito Tanaka, Koichiro Sato, Teppei Omori, Kunihiro Takanashi, Yoshikazu Hayashi, Yuki Nakajima, Yasuyuki Miyakura, Takayuki Matsumoto, Naohisa Yoshida, Motohiro Esaki, Toshio Uraoka, Hiroyuki Kato, Yuji Inoue, Boyuan Peng, Ruiyao Zhang, Takashi Hisabe, Tomoki Matsuda, Hironori Yamamoto, Noriko Tanaka, Alan Kawarai Lefor, Xin Zhu, Kazutomo Togashi
来源：
GASTROINTESTINAL ENDOSCOPY
摘要：
结直肠癌深层黏膜下侵袭（T1b）与浅层侵袭（T1a）或不侵袭（Tis）的区分并不简单。本研究旨在开发一种计算机辅助诊断系统（CADx），仅利用非放大内窥镜白光图像来建立早期癌症的诊断。共收集了1470名患者的1513个病变（Tis 1074，T1a 145，T1b 294）在5108个图像中，并将其分配为训练和测试数据集（3：1）。使用ResNet-50网络作为骨干提取图像特征。使用过采样和聚焦损失来补偿侵袭阶段的类别不平衡。使用包括403个结肠直肠癌和1392个图像的测试数据集来评估诊断性能。两位专家和两位实习医生读取相同的测试数据集。在每个病变分数的90％截断处，CADx显示出最高的特异性94.4％[95％置信区间：91.3-96.6]，敏感性为59.8％[48.3-70.4]，准确性为87.3％[83.7-90.4]。特征曲线下的面积为CADx为85.1％[79.9-90.4]，专家1为88.2％[83.7-92.8]，专家2为85.9％[80.9-90.9]，实习医生1为77.0％[71.5-82.4]（与CADx比较：p = 0.0076），实习医生2为66.2％[60.6-71.9]（P <0.0001）。该功能还在九个短视频中得到确认。使用内窥镜白光图像开发的CADx显示出对T1b病变诊断的优秀病变特异性和准确性，与专家相当且优于实习医生。（UMIN000037053）（249<=250个字）。Copyright © 2023 American Society for Gastrointestinal Endoscopy. Published by Elsevier Inc. All rights reserved.
Differentiation of colorectal cancers with deep submucosal invasion (T1b) from colorectal cancers with superficial invasion (T1a) or no invasion (Tis) is not straightforward. This study aimed to develop a computer aided diagnosis system (CADx) to establish the diagnosis of early-stage cancers using non-magnified endoscopic white light images alone.A total of 1513 lesions (Tis 1074, T1a 145, T1b 294) in 5108 images were collected from 1470 patients at ten academic hospitals and assigned to training and testing datasets (3:1). The ResNet-50 network was used as the backbone to extract features from images. Over sampling and focal loss were used to compensate class imbalance of invasive stage. Diagnostic performance was assessed using the testing dataset including 403 CRCs with 1392 images. Two experts and two trainees read the identical testing dataset.At 90% cutoff for per lesion score, CADx showed the highest specificity of 94.4% [95% confidence interval: 91.3 - 96.6], with 59.8% [48.3 - 70.4] sensitivity and 87.3% [83.7 - 90.4] accuracy. The area under the characteristic curve was 85.1% [79.9 - 90.4] for CADx, 88.2% [83.7 - 92.8] for expert 1, 85.9% [80.9 - 90.9] for expert 2, 77.0% [71.5 - 82.4] for trainee 1 (vs. CADx: p=0.0076), and 66.2% [60.6 - 71.9] for trainee 2 (p<0.0001). The function was also confirmed on nine short videos.CADx developed with endoscopic white light images showed excellent per lesion specificity and accuracy for T1b lesion diagnosis, equivalent to experts and superior to trainees. (UMIN000037053) (249 =<250 words).Copyright © 2023 American Society for Gastrointestinal Endoscopy. Published by Elsevier Inc. All rights reserved.