BACKGROUND

Artificial intelligence (AI) is increasingly being used to detect neoplasia and interpret endoscopic images. The T stage of Barrett's carcinoma is a major criterion for subsequent treatment decisions. Although endoscopic ultrasound is still the standard for preoperative staging, its value is debatable. Novel tools are required to assist with staging, to optimize results. This study aimed to investigate the accuracy of T stage of Barrett's carcinoma by an AI system based on endoscopic images.

METHODS

1020 images (minimum one per patient, maximum three) from 577 patients with Barrett's adenocarcinoma were used for training and internal validation of a convolutional neural network. In all, 821 images were selected to train the model and 199 images were used for validation.

RESULTS

AI recognized Barrett's mucosa without neoplasia with an accuracy of 85 % (95 %CI 82.7-87.1). Mucosal cancer was identified with a sensitivity of 72 % (95 %CI 67.5-76.4), specificity of 64 % (95 %CI 60.0-68.4), and accuracy of 68 % (95 %CI 64.6-70.7). The sensitivity, specificity, and accuracy for early Barrett's neoplasia < T1b sm2 were 57 % (95 %CI 51.8-61.0), 77 % (95 %CI 72.3-80.2), and 67 % (95 %CI 63.4-69.5), respectively. More advanced stages (T3/T4) were diagnosed correctly with a sensitivity of 71 % (95 %CI 65.1-76.7) and specificity of 73 % (95 %CI 69.7-76.5). The overall accuracy was 73 % (95 %CI 69.6-75.5).

CONCLUSIONS

The AI system identified esophageal cancer with high accuracy, suggesting its potential to assist endoscopists in clinical decision making.