A Multicenter Pivotal Study on the Artificial Intelligence System for Neoplastic Lesions Detection in Upper Gastrointestinal Endoscopy.

This pivotal study aimed to evaluate the performance of the CAD-EYE prototype in identifying esophageal squamous cell carcinoma (ESCC) and gastric neoplasm (GN) for regulatory approval of the Pharmaceuticals and Medical Devices Agency in Japan. This retrospective study utilized image datasets comprising 15 consecutive video frames captured using non-magnifying white-light imaging (WLI), blue laser/light imaging (BLI), and linked color imaging (LCI). The sensitivity and specificity of the CAD-EYE prototype for successful detection were calculated using the gold standard, which consists of image datasets of neoplastic lesions annotated by experienced endoscopists. A total of 620, 679, and 682 ESCC datasets were analyzed in the WLI, BLI, and LCI groups, respectively. The sensitivity and specificity of ESCC detection were 85.9% 81.1%-90.6% and 93.3% 90.8%-95.7% in the WLI group, 97.6% 95.6%-99.7% and 92.9% 90.6%-95.3% in the BLI group, and 96.6% 94.2%-99.1% and 93.2% 91.0%-95.5% in the LCI group. The sensitivities for pT1a ESCC were 85.3%, 97.3%, and 97.2% in the WLI, BLI, and LCI groups, respectively. For GN, 841 WLI and 882 LCI datasets were analyzed. The sensitivity, specificity, and specificity in the detection flag of GN detection were 95.5% 92.8%-98.1%, 86.1%, and 85.4% 82.6%-88.2% in the WLI group, and 93.9% 91.1%-96.7%, 94.4%, and 93.9% 92.0%-95.8% in the LCI group, respectively. The sensitivities for pT1a early gastric cancer were 93.8% and 92.4% in the WLI and LCI groups, respectively. The CAD-EYE prototype demonstrated high sensitivity in detecting ESCC and GN, highlighting its potential as a promising tool for clinical applications.