A computer-aided diagnosis system using white-light endoscopy for the prediction of conventional adenoma with high grade dysplasia

Abstract

Objectives: We developed a computer-aided diagnosis system called ECRCCAD using standard white-light endoscopy (WLE) for predicting conventional adenomas with high-grade dysplasia (HGD) to optimise the patients' management decisions during colonoscopy.

Methods: Pretraining model was used to fine-tune the model parameters by transfer learning. 2,397 images of HGD and 2,487 low-grade dysplasia (LGD) images were randomly assigned (8:1:1) to the training, optimising, and internal validation dataset. The prospective validation dataset is the frames accessed from colonoscope videoes. One independent rural hospital provided an external validation dataset. Histopathological diagnosis was used as the standard criterion. The capability of the ECRCCAD to distinguish HGD was assessed and compared with two expert endoscopists.

Results: The accuracy, sensitivity and specificity for diagnosis of HGD in the internal validation set were 90.5%, 93.2%, 87.9%, respectively. While 88.2%, 85.4%, 89.8%, respectively, for the external validation set. For the prospective validation set, ECRCCAD achieved an AUC of 93.5% in diagnosing HGD. The performance of ECRCCAD in diagnosing HGD was better than that of the expert endoscopist in the external validation set (88.2% vs. 71.5%, P < 0.0001).

Conclusion: ECRCCAD had good diagnostic capability for HGD and enabled a more convenient and accurate diagnosis using WLE.

Keywords: Artificial intelligence; Colorectal cancer; Computer-aided diagnosis system; High grade dysplasia; White light endoscopy.