A Novel Deep Learning System for Diagnosing Early Esophageal Squamous Cell Carcinoma: A Multicenter Diagnostic Study.

Introduction: This study aims to construct a real-time deep convolutional neural networks (DCNNs) system to diagnose early esophageal squamous cell carcinoma (ESCC) with white light imaging endoscopy.

Methods: A total of 4,002 images from 1,078 patients were used to train and cross-validate the DCNN model for diagnosing early ESCC. The performance of the model was further tested with independent internal and external validation data sets containing 1,033 images from 243 patients. The performance of the model was then compared with endoscopists. The accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and Cohen kappa coefficient were measured to assess performance.

Results: The DCNN model had excellent performance in diagnosing early ESCC with a sensitivity of 0.979, a specificity of 0.886, a positive predictive value of 0.777, a negative predictive value of 0.991, and an area under curve of 0.954 in the internal validation data set. The model also depicted a tremendously generalized performance in 2 external data sets and exhibited superior performance compared with endoscopists. The performance of the endoscopists was markedly elevated after referring to the predictions of the DCNN model. An open-accessed website of the DCNN system was established to facilitate associated research.

Discussion: A real-time DCNN system, which was constructed to diagnose early ESCC, showed good performance in validation data sets. However, more prospective validation is needed to understand its true clinical significance in the real world.