ROI-guided virtual narrow band imaging for laryngeal cancer screening

White light laryngoscopy is widely available but can miss subtle vascular changes associated with early laryngeal neoplasia. We developed a region-of-interest (ROI)-guided cycle-consistent generative adversarial network (CycleGAN) that translates white light imaging (WLI) images into virtual narrow band imaging (NBI) images, while emphasizing expert-annotated lesion regions. The framework was trained and evaluated on 775 weakly paired WLI-NBI image pairs and assessed with technical image quality metrics, subjective reader scoring, a visual Turing test, and a multi-center reader study involving 12 otolaryngologists from nine institutions. The model preserved lesion-region structure better than global background structure and improved overall reader accuracy compared with WLI alone (81.5% versus 73.4%). Assistance mainly increased sensitivity for junior readers and specificity for senior readers. These findings support virtual NBI as a potential software-based adjunct where optical NBI is unavailable, while requiring further external and small-lesion validation.