Validation of Artificial Intelligence Computer-Aided Detection on Gastric Neoplasm in Upper Gastrointestinal Endoscopy

Background/objectives: Gastric cancer ranks fifth for incidence and fourth in the leading causes of mortality worldwide. In this study, we aimed to validate previously developed artificial intelligence (AI) computer-aided detection (CADe) algorithm, called ALPHAON in detecting gastric neoplasm.

Methods: We used the retrospective data of 500 still images, including 5 benign gastric ulcers, 95 with gastric cancer, and 400 normal images. Thereby we validated the CADe algorithm measuring accuracy, sensitivity, and specificity with the result of receiver operating characteristic curves (ROC) and area under curve (AUC) in addition to comparing the diagnostic performance status of four expert endoscopists, four trainees, and four beginners from two university-affiliated hospitals with CADe algorithm. After a washing-out period of over 2 weeks, endoscopists performed gastric detection on the same dataset of the 500 endoscopic images again marked by ALPHAON.

Results: The CADe algorithm presented high validity in detecting gastric neoplasm with accuracy (0.88, 95% CI: 0.85 to 0.91), sensitivity (0.93, 95% CI: 0.88 to 0.98), specificity (0.87, 95% CI: 0.84 to 0.90), and AUC (0.962). After a washing-out period of over 2 weeks, overall validity improved in the trainee and beginner groups with the assistance of ALPHAON. Significant improvement was present, especially in the beginner group (accuracy 0.94 (0.93 to 0.96) < 0.001, sensitivity 0.87 (0.82 to 0.92) < 0.001, specificity 0.96 (0.95 to 0.97) < 0.001).

Conclusions: The high validation performance state of the CADe algorithm system was verified. Also, ALPHAON has demonstrated its potential to serve as an endoscopic educator for beginners improving and making progress in sensitivity and specificity.