Feed

Generative indigo carmine chromoendoscopy (generative chromoendoscopy), produced using deep learning-based synthetic image transformation, is a novel image-enhanced endoscopic approach whose clinical feasibility has not been validated. This proof-of-concept study evaluated whether generative chromoendoscopy improves the visibility of gastric neoplasms compared with conventional white-light imaging (WLI) and real chromoendoscopy.

This prospective, multicenter study was conducted at three Japanese institutions between January and August 2025. A deep neural network-based program converted WLI into generative chromoendoscopy images. Patients with early gastric cancer or adenoma sequentially underwent endoscopy using WLI, generative chromoendoscopy, and real chromoendoscopy. During live endoscopy, two endoscopists independently assessed lesion visibility using a 7-point Likert-type scale (-3 to +3). The primary endpoint was visibility of gastric neoplasms with generative chromoendoscopy relative to WLI, with comparison to real chromoendoscopy. Secondary endpoints included clinicopathologic factors associated with improved visibility.

Sixty patients were enrolled. Mean visibility scores were 0.67 ± 0.93 for generative chromoendoscopy and 0.51 ± 1.00 for real chromoendoscopy. The mean difference was 0.16 (95% confidence interval [CI] 0.01-0.30). Differentiated-type histology (odds ratio [OR] 2.19, 95% CI 1.03-4.62), current Helicobacter pylori infection (OR 2.14, 95% CI 1.02-4.46), and expert endoscopist status (OR 2.14, 95% CI 1.22-3.76) were significantly associated with higher visibility scores using generative chromoendoscopy.

Generative chromoendoscopy was feasible during live endoscopic procedures. It improved visibility of pre-identified gastric neoplasms compared with WLI and achieved visibility comparable to real chromoendoscopy.