249. AUTOMATED PREDICTION SYSTEM USING COMPUTER VISION FOR ESOPHAGEAL STRICTURE AFTER ESOPHAGEAL ENDOSCOPIC RESECTION

Background

Although endoscopic submucosal dissection (ESD) is a standard treatment for early esophageal cancer, post-treatment esophageal stricture (ES) is the most common complication, especially in patients with circumferential resection range greater than three-quarters. This study sought to establish an automated system for predicting ES after ESD from endoscopic video using computer vision. We believe that this automation can help endoscopists to perform appropriate decision-making, such as whether to perform steroid injection during ESD or preventative balloon dilatation after ESD.

Methods

This study assessed recorded endoscopic videos of 73 patients who underwent ESD for superficial esophageal cancer. A deep neural network-based classifier—DeepES—was developed to assess whether ES would occur. The model was trained and tested using k-4 folds cross-validation, and the area under the curve (AUC) values of DeepES were calculated to validate the model performance against several clinicopathological factors.

Results

For the 73 patients evaluated by k-4 folds cross-validation, overall accuracy for the DeepES was 0.712, meaning ES could be correctly predicted by AI for 52 patients.

The AUC value of DeepES was higher (AUC 0.709) compared with circumferential resection range (AUC 0.680) and pathological tumor depth (AUC 0.570).

Conclusion

DeepES has a higher accuracy for predicting ES during ESD than other factors, including circumferential resection range and pathological tumor depth. Use of DeepES enables optimal decision-making during ESD, such as whether to perform preventative strategies, including balloon dilatation. This report is the first to show the value of state-of-the-art technologies using computer vision for ESD to prevent ES.