Evaluation of computer-aided detection for gastric cancer using white-light and linked-color imaging: a pilot study

Background and Aims: In recent years, the field of endoscopic artificial intelligence has seen significant advancements, largely because of the widespread implementation of deep learning techniques. However, the computer-aided detection (CADe) of the stomach poses significant challenges in clinical practice. Here, we evaluated the performance of a newly developed CADe system, CAD EYE (Fujifilm, Tokyo, Japan), by comparing the frequency of the detection box appearance with white-light imaging (WLI) versus linked-color imaging (LCI) during the process of detecting gastric cancer (GC) and detection of GC with and without CADe. Methods: This single-center observational retrospective study included 105 patients who underwent esophagogastroduodenoscopy (EGD) using CADe and 105 controls selected by propensity-score matching from 600 patients. The primary outcome was to compare the detection box appearance of WLI and LCI during the CADe observation. Secondary outcomes included comparisons of biopsy rates, examination times, and cancer detection rates between groups. Furthermore, we investigated whether the landmark checker could accurately identify the stomach site. Biopsy rates and examination times were comparable between the groups. CADe accurately detected all 18 known-early GC cases. The landmark checker function identified an average of 5.72 of 7 key gastric sites (81.7%). Conclusions: This pilot study suggests that CADe, particularly when combined with LCI, may enhance GC detection during EGD without significantly increasing the examination time. Although promising, a high false-positive rate indicates that further optimization is needed.