Diagnostic evaluation of a deep learning model for optical diagnosis of colorectal cancer

Abstract

Colonoscopy is commonly used to screen for colorectal cancer (CRC). We develop a deep learning model called CRCNet for optical diagnosis of CRC by training on 464,105 images from 12,179 patients and test its performance on 2263 patients from three independent datasets. At the patient-level, CRCNet achieves an area under the precision-recall curve (AUPRC) of 0.882 (95% CI: 0.828–0.931), 0.874 (0.820–0.926) and 0.867 (0.795–0.923). CRCNet exceeds average endoscopists performance on recall rate across two test sets (91.3% versus 83.8%; two-sided t-test, p < 0.001 and 96.5% versus 90.3%; p = 0.006) and precision for one test set (93.7% versus 83.8%; p = 0.02), while obtains comparable recall rate on one test set and precision on the other two. At the image-level, CRCNet achieves an AUPRC of 0.990 (0.987–0.993), 0.991 (0.987–0.995), and 0.997 (0.995–0.999). Our study warrants further investigation of CRCNet by prospective clinical trials.