Efficacy of Digestive Endoscope Based on Artificial Intelligence System in Diagnosing Early Esophageal Carcinoma.

Objective: To explore the efficacy of digestive endoscopy (DEN) based on artificial intelligence (AI) system in diagnosing early esophageal carcinoma.

Methods: The clinical data of 300 patients with suspected esophageal carcinoma treated in our hospital from January 2018 to January 2020 were retrospectively analyzed; among them, 198 were diagnosed with esophageal carcinoma after pathological examination, and 102 had benign esophageal lesion. An AI system based on convolutional neural network (CNN) was adopted to assess the DEN images of patients with early esophageal carcinoma. A total of 200 patients (148 with early esophageal carcinoma and 52 with benign esophageal lesion) were selected as the learning group for the Inception V3 image classification system to learn; and the rest 100 patients (50 with early esophageal carcinoma and 50 with benign esophageal lesion) were included in the diagnosis group for the Inception V3 system to assist the narrow-band imaging (NBI) with diagnosis. The diagnosis results from Inception V3-assisted NBI were compared with those from imaging physicians, and the diagnostic efficacy diagram was drawn.

Results: The diagnosis rate of AI-NBI was significantly faster than that of physician diagnosis (0.02 ± 0.01 vs. 5.65 ± 0.32 s (mean rate of two physicians), < 0.001); between AI-NBI diagnosis and physician diagnosis, no statistical differences in sensitivity (90.0% vs. 92.0%), specificity (92.0% vs. 94.0%), and accuracy (91.0% vs. 93.0%) were observed ( > 0.05); and according to the ROC curves, AUC (95% CI) of AI-NBI diagnosis = 0.910 (0.845-0.975), and AUC (95% CI) of physician diagnosis = 0.930 (0.872-0.988).

Conclusion: CNN-based AI system can assist NBI in screening early esophageal carcinoma, which has a good application prospect in the clinical diagnosis of early esophageal carcinoma.