Computer-aided diagnosis of early-stage colorectal cancer using nonmagnified endoscopic white-light images (with videos)
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
DOI: https://doi.org/10.1016/j.gie.2023.01.050
Abstract:Background and aims: Differentiation of colorectal cancers (CRCs) with deep submucosal invasion (T1b) from CRCs with superficial invasion (T1a) or no invasion (Tis) is not straightforward. This study aimed to develop a computer-aided diagnosis (CADx) system to establish the diagnosis of early-stage cancers using nonmagnified endoscopic white-light images alone. Methods: From 5108 images, 1513 lesions (Tis, 1074; T1a, 145; T1b, 294) were collected from 1470 patients at 10 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. Oversampling and focal loss were used to compensate class imbalance of the invasive stage. Diagnostic performance was assessed using the testing dataset including 403 CRCs with 1392 images. Two experts and 2 trainees read the identical testing dataset. Results: At a 90% cutoff for the per-lesion score, CADx showed the highest specificity of 94.4% (95% confidence interval [CI], 91.3-96.6), with 59.8% (95% CI, 48.3-70.4) sensitivity and 87.3% (95% CI, 83.7-90.4) accuracy. The area under the characteristic curve was 85.1% (95% CI, 79.9-90.4) for CADx, 88.2% (95% CI, 83.7-92.8) for expert 1, 85.9% (95% CI, 80.9-90.9) for expert 2, 77.0% (95% CI, 71.5-82.4) for trainee 1 (vs CADx; P = .0076), and 66.2% (95% CI, 60.6-71.9) for trainee 2 (P < .0001). The function was also confirmed on 9 short videos. Conclusions: A CADx system developed with endoscopic white-light images showed excellent per-lesion specificity and accuracy for T1b lesion diagnosis, equivalent to experts and superior to trainees. (Clinical trial registration number: UMIN000037053.).