Abstract
BACKGROUND AND STUDY AIMS: Indigo carmine chromoendoscopy (IC) enhances diagnosis of early gastric cancer (EGC), but its clinical application is limited by procedure complexity and time. We developed a deep-learning system using a cycle-consistent generative adversarial network (CycleGAN) to generate virtual IC images from white-light endoscopy (WLE) and evaluated visibility of EGC in video-based virtual IC in a pilot study. PATIENTS AND METHODS: We collected 4,096 endoscopic still images (2,089 WLE, 2,007 real IC) from 262 patients with gastric neoplasms. A CycleGAN model was trained to convert WLE into virtual IC images, and videos with 512 × 512 pixels at 30 frames per second were generated for five EGC cases. For each case, WLE, real IC, and virtual IC videos were prepared and evaluated by 16 endoscopists (6 experts, 10 non-experts). Visibility relative to WLE was rated using a 7-point Likert-type scale (-3 to +3), with positive values indicating improved visibility. RESULTS: A total of 160 evaluations were performed. Median [IQR] visibility score was 1 [0-2)] for real IC and 0 [-1 to 1] for virtual IC ( P < 0.001). In virtual IC, 46.3% of cases achieved a score of +1 or higher. Scores significantly varied by endoscope system ( P < 0.001). CONCLUSIONS: Virtual IC improved visibility compared with WLE in nearly half the assessments, although its efficacy did not equal real IC. Optimizing performance for specific endoscope systems may enhance its clinical utility as a practical alternative for improving EGC detection.