AI Assisted Endoscopy for Gastric Cancer Diagnosis: The First Japanese Clinical Trial

Speaker

Mitsuru KAISE Professor Department of Gastroenterology Nippon Medical School, Japan

Mitsuru KAISE
Professor
Department of Gastroenterology
Nippon Medical School, Japan

Talk Abstract

Background and aims: The first multicenter, randomized, control trial was conducted to evaluate whether AI assistance improves the diagnostic performance of superficial gastric neoplasia.

Methods: Patients undergoing endoscopy prior to endoscopic resection (ER) of gastric neoplasia or for surveillance after ER were randomized (1:1) to the AI-assisted and non-AI-assisted groups. Endoscopy was performed by a non-expert endoscopist blinded to the patient information. The AI system provides a confidence level for gastric neoplasia on a still image captured during endoscopy and indicates neoplasia if the level = or > a 60% cut-off. The target lesions were those diagnosed as gastric neoplasia and those diagnosed as non-neoplasia but requiring biopsy. The primary outcome was non-expert diagnostic accuracy.

Results: 312 patients with 265 gastric neoplasias and 164 non-neoplasias were enrolled. Non-expert diagnostic accuracy was 65.3% vs. 59.9% (P = .24), in the AI-assisted and non-AI-assisted groups, respectively. AI and non-AI endoscopists' diagnoses differed in 21% of cases, but AI assistance changed the endoscopists' final diagnosis in 4% of cases, all of which were pathologically correct. The receiver operating characteristic analysis showed that the highest the confidence level cut-off was 78.5%, despite the current setting of 60%. Flat or superficially depressed small neoplasias tended to be missed by the AI.

Conclusion: The present AI assistance provided a small positive effect in the diagnosis of superficial gastric neoplasia for non-experts. Further improvements in gastric AI are needed by changing the confidence level cut-off and additionally training the AI on easily missed neoplasias.