Artificial Intelligence Similar to Expert Endoscopists at Differentiating Between Mucosal and Submucosal Barrett’s Cancer

Artificial intelligence (AI) systems using convolutional neural networks have been developed to differentiate between esophageal cancer (EAC), high-grade dysplasia, and nondysplastic Barrett’s esophagus. However, until now, the ability to differentiate mucosal from submucosal invasion has not been evaluated. This differentiation is important because it can be used to determine endoscopic treatment modalities and influence prognostication.

This retrospective, multicenter study used white-light endoscopy images from patients with histologically confirmed T1a and T1b EAC at 3 tertiary care centers in Germany. Different images were used for training and validation of the AI system. Overall, 108 images from mucosal (T1a) lesions and 122 images from submucosal (T1b) lesions were utilized. The AI system was trained with still images of various sizes and resized to 512 x 512 pixels for uniform resolution. The performance of the AI system was compared to that of expert endoscopists. Five international experts, blinded to the diagnosis of the images, were asked to characterize the still endoscopic images for comparison. The performance of the AI system was calculated over 10 runs. 

The sensitivity and specificity of the AI network to distinguish T1a versus T1b BC were 0.77 (95% confidence interval [CI], 0.75-0.78) and 0.64 (95% CI, 0.62-0.66), respectively. The accuracy of the system was calculated at 0.71 (95% CI, 0.70-0.72). With the experts, sensitivity was 0.63 (95% CI, 0.53-0.78), specificity was 0.78 (95% CI, 0.67-0.89), and accuracy was 0.70 (95% CI, 0.67-0.73). Multivariate analysis using the McNemar test found no significant difference between the accuracy of the AI system and the experts.