Diffusion-based image translation from white light to narrow-band imaging in gastrointestinal endoscopy

Abstract

Narrow-band imaging (NBI) enhances vascular and mucosal visualization, enabling early detection of gastrointestinal lesions. However, its adoption is limited by hardware constraints and costs, leaving white light endoscopy (WLE) as the widely used but diagnostically inferior modality. Translating WLE into realistic NBI-like images provides a scalable solution to improve diagnostic workflows, generate synthetic datasets, and facilitate multi-modality analysis. Translating WLE images into realistic NBI-like images is challenging due to the lack of paired WLE-NBI image datasets for training and the complex, varied nature of lesions in gastrointestinal endoscopy, which often involve rich details and subtle textures. In this study, we propose a novel diffusion-based framework tailored for WLE-to-NBI image translation. Leveraging stable diffusion with domain-specific enhancements, our method integrates LoRA fine-tuning to embed NBI-specific features and employs a self-attention injection mechanism to dynamically incorporate vascular and mucosal patterns while preserving the spatial structure and semantic integrity of the input WLE images. This approach ensures fine-grained feature translation and structural fidelity crucial for medical applications. Quantitative and qualitative experiments highlight the superiority of the proposed approach in generating high-fidelity NBI-like images. Furthermore, it demonstrates potential for data augmentation and robustness in long-range video frame registration, offering a reliable solution for enhancing clinical decision-making.