Introduction

Early detection of gastric neoplasms in esophagogastroduodenoscopy (EGD) remains challenging due to the subtle and indistinct appearance of the lesions. In addition to white light imaging (WLI), image-enhanced endoscopy techniques, such as linked color imaging (LCI), are commonly used to detect such lesions. Recently, a computer-aided detection (CADe) system compatible with WLI and LCI (CAD EYE; Fujifilm Co., Tokyo, Japan) has been developed. However, evidence of its effectiveness in real-world clinical settings remains limited.

Aims & Methods

We conducted a multicenter, non-randomized confirmatory trial to evaluate the performance of CAD EYE in detecting gastric neoplasms. Between July 2023 and January 2024, patients undergoing EGD for known early gastric neoplasms were recruited at four institutions in Japan. While systematically observing the entire stomach with WLI and LCI, we investigated whether these lesions were detected by CAD EYE. The primary outcome was the sensitivity of CAD EYE (WLI-CADe and LCI-CADe) in detecting tumors. Secondary outcomes included identifying tumor characteristics and factors associated with missed detections by CAD EYE. To determine the sample size, we presumed the sensitivity of CAD EYE to be 95% based on pre-approval data, with a threshold set at 90%. The number of lesions required was calculated to be 150, ensuring that the lower limit of the 95% confidence interval (CI) exceeded the threshold of 90%. Assuming one lesion per patient and accounting for a 5% dropout rate, we recruited a total of 160 patients.

Results

A total of 174 lesions from 162 patients (mean age: 73.7±9.5 years, male: 69.7%) were analyzed. Current or past history of Helicobacter pylori (H. pylori) infection was observed in 158 patients (97.5%). The mean tumor diameter was 15.1 ± 11.3 mm, and the morphology consisted of 6 protruded, 51 flat elevated, and 117 flat types. Histology included 151 differentiated adenocarcinoma, 20 undifferentiated adenocarcinoma, 2 adenocarcinoma of fundic-gland type, and 1 tubular adenoma. The sensitivities of WLI-CADe and LCI-CADe in detecting tumors were 92.0% (95% CI: 86.9%–95.5%) and 96.5% (95% CI: 92.6%–98.7%), respectively. The sensitivity of LCI-CADe met expectations, as the lower limit of the 95% CI exceeded the threshold of 90%; however, WLI-CADe failed to meet the same standard. In WLI-CADe, lesions located on the anterior/posterior wall were missed significantly more often than those on the greater/lesser curvature (13.9% [11 of 79] vs. 3.2% [3 of 95]; P = 0.020); whereas no difference was observed in LCI-CADe. Other factors, such as H. pylori infection status, tumor morphology, location, and histology, showed no differences between lesions detected and missed by WLI-CADe and LCI-CADe.

Conclusion

The use of CADe with LCI demonstrated satisfactory sensitivity in detecting gastric neoplasms in real-world clinical settings, suggesting its potential to assist endoscopists in tumor detection. However, CADe with WLI did not meet the expected standards, possibly due to limitations, such as detecting lesions located on the anterior/posterior wall.

Disclosure

CAD EYE was temporarily provided by Fujifilm Co. Osamu Dohi has received a research grant, not related to this study, from Fujifilm Co. The remaining authors declare no conflicts of interest.