Artificial intelligence-based virtual iodine staining model for esophageal squamous cell carcinoma: a multicenter proof-of-concept study.

BACKGROUND

Early detection of esophageal squamous cell carcinoma (ESCC) is critical, but Lugol chromoendoscopy (LCE) has limitations including patient discomfort and technical difficulty. Artificial intelligence (AI)-based virtual staining offers a potential solution. This study aimed to develop and evaluate multiple AI models for converting esophageal white light images into virtual iodine-stained images.

METHODS

Based on three mainstream staining algorithms, seven AI models were trained and validated using retrospectively collected white light and real iodine-stained images from one center and tested on independent data from another center. Objective metrics (SSIM, ΔCF, FID, and CMMD) and subjective evaluations by ten endoscopists (lesion visibility, boundary clarity, and real‑versus‑virtual discrimination) were used to select the best‑performing model.

RESULTS

The training and validation datasets comprised 7209 and 802 images, respectively; the test dataset contained 439 images. Among the seven models, CUT achieved the best overall objective scores (FID: 42.21, ΔCF: 1.70, CMMD: 0.190 on validation; and 59.08, 2.03, 0.460 on test) and was rated highest by endoscopists for lesion visibility (mean score 3.27 ± 0.05 on a 5‑point scale).

CONCLUSIONS

The virtual iodine-stained images generated by the CUT model provide superior lesion visibility and favorable feature distribution for ESCC. This pilot proof‑of‑concept study suggests that the CUT model holds promise as a potential alternative to conventional LCE and may offer a way to reduce patient discomfort. Its applicability warrants further investigation in larger prospective cohorts.