Artificial intelligence-based marks detection and incision guide line prediction model in esophageal endoscopic submucosal dissection: a multicenter study (with video).

BACKGROUND

Endoscopic submucosal dissection (ESD) was an important minimally invasive procedure for treating early esophageal cancer, where the surrounding mucosal incision (SMI) was a crucial yet challenging phase. This study aimed to develop an artificial intelligence (AI) model for marks detection and incision guide line prediction during SMI, and to evaluate its performance compared to multicenter endoscopists.

METHODS

Images were extracted at one frame per second from video clips and divided into 3 datasets: training dataset, internal and external test dataset. Twenty-two endoscopists participated in the comparison with the model, including eight senior and fourteen junior endoscopists. Multiple objective indicators and a marks-guide line subjective score (M-GSS) were used for outcomes.

RESULTS

A total of 46,280 SMI images were extracted from 166 esophageal ESD videos in 4 hospitals between 2016 and 2024. For marks detection, the precision of the AI model in the training dataset, internal, and external test datasets were 78.97%, 85.76%, and 79.46%. For the incision guide line, the average distance error (ADE) was 0.096, 0.097, and 0.159, respectively. The AI model achieved an accuracy (63.21% vs 54.57%, P < 0.001), ADE (0.081 vs 0.241, P = 0.002), and M-GSS (1.88 vs 1.75, P < 0.001) significantly better than that of junior endoscopists, and were comparable with those of senior endoscopists.

CONCLUSIONS

The AI model achieved promising results on large SMI image datasets, showing comparable effectiveness with senior endoscopists for marks detection and incision guide line, which had the potential to navigate SMI safely and properly.