Diagnostic performance of artificial intelligence to identify deeply invasive colorectal cancer on non-magnified plain endoscopic images

Abstract

Background and study aims Colorectal cancers (CRC) with deep submucosal invasion (T1b) could be metastatic lesions. However, endoscopic images of T1b CRC resemble those of mucosal CRCs (Tis) or with superficial invasion (T1a). The aim of this study was to develop an automatic computer-aided diagnosis (CAD) system to identify T1b CRC based on plain endoscopic images. Patients and methods In two hospitals, 1839 non-magnified plain endoscopic images from 313 CRCs (Tis 134, T1a 46, T1b 56, beyond T1b 37) with sessile morphology were extracted for training. A CAD system was trained with the data augmented by rotation, saturation, resizing and exposure adjustment. Diagnostic performance was assessed using another dataset including 44 CRCs (Tis 23, T1b 21) from a third hospital. CAD generated a probability level for T1b diagnosis for each image, and > 95 % of probability level was defined as T1b. Lesions with at least one image with a probability level > 0.95 were regarded as T1b. Primary outcome is specificity. Six physicians separately read the same testing dataset. Results Specificity was 87 % (95 % confidence interval: 66–97) for CAD, 100 % (85–100) for Expert 1, 96 % (78–100) for Expert 2, 61 % (39–80) for both gastroenterology trainees, 48 % (27–69) for Novice 1 and 22 % (7–44) for Novice 2. Significant differences were observed between CAD and both novices (P = 0.013, P = 0.0003). Other diagnostic values of CAD were slightly lower than of the two experts. Conclusions Specificity of CAD was superior to novices and possibly to gastroenterology trainees but slightly inferior to experts.