Deep-Learning for the Diagnosis of Esophageal Cancers and Precursor Lesions in Endoscopic Images: A Model Establishment and Nationwide Multicenter Performance Verification Study

Abstract

Background: Suspicion of lesions and prediction of the histology of esophageal cancers or premalignant lesions in endoscopic images are not yet accurate. The local feature selection and optimization functions of the model enabled an accurate analysis of images in deep learning. Objectives: To establish a deep-learning model to diagnose esophageal cancers, precursor lesions, and non-neoplasms using endoscopic images. Additionally, a nationwide prospective multicenter performance verification was conducted to confirm the possibility of real-clinic application. Methods: A total of 5162 white-light endoscopic images were used for the training and internal test of the model classifying esophageal cancers, dysplasias, and non-neoplasms. A no-code deep-learning tool was used for the establishment of the deep-learning model. Prospective multicenter external tests using 836 novel images from five hospitals were conducted. The primary performance metric was the external-test accuracy. An attention map was generated and analyzed to gain the explainability. Results: The established model reached 95.6% (95% confidence interval: 94.2–97.0%) internal-test accuracy (precision: 78.0%, recall: 93.9%, F1 score: 85.2%). Regarding the external tests, the accuracy ranged from 90.0% to 95.8% (overall accuracy: 93.9%). There was no statistical difference in the number of correctly identified the region of interest for the external tests between the expert endoscopist and the established model using attention map analysis (P = 0.11). In terms of the dysplasia subgroup, the number of correctly identified regions of interest was higher in the deep-learning model than in the endoscopist group, although statistically insignificant (P = 0.48). Conclusions: We established a deep-learning model that accurately classifies esophageal cancers, precursor lesions, and non-neoplasms. This model confirmed the potential for generalizability through multicenter external tests and explainability through the attention map analysis.