Structure-Aware Local Affine CycleGAN for Balanced and High-Contrast Enhancement in Virtual Narrow-Band Imaging

Objective: To investigate a method for converting unpaired white-light endoscopic images (white-light imaging, WLI) into a narrow-band imaging (NBI)-like style and, on this basis, enhancing local contrast of vessel-like structures. Methods: We constructed a constrained local-affine CycleGAN that performs image-style translation through spatially varying RGB affine transformations. We then estimated an enhancement-support map from nine features extracted from source WLI and clinical regional annotations, and used a unified RGB attenuation operator to generate a balanced mode. The high-contrast mode further incorporated a spatial network and a learned fusion gate. Two local-affine bases with different random seeds were each trained for 20,000 updates. After the enhancement module was frozen, the outputs were compared with the input, standard CycleGAN and CUT on six PolypsSet development images containing radius-12 local vessel–background pairs. The primary metrics were signed contrast-to-noise ratio (CNR) and Weber contrast. Results: Across the two base repeats, the balanced mode achieved mean CNRs of 3.380–3.430 and a Weber contrast of 0.262, compared with 3.387 and 0.173 for standard CycleGAN, respectively. Across the four base–enhancement combinations, the high-contrast mode achieved mean CNRs of 5.954–6.141 and Weber contrasts of 0.337–0.342. In per-image leave-one-out analyses, the differences in CNR and Weber contrast between each high-contrast combination and standard CycleGAN or CUT remained positive. The mean additional luminance attenuation in confirmed-background regions was 0.01398–0.02009 for high-contrast mode and approximately 0.025 for balanced mode. Conclusions: In this set of six development images and their fixed local annotations, staged translation and support-map enhancement increased local vessel–background contrast; the high-contrast mode produced larger numerical contrast gains. Responses from local dark networks and background textures indicate that contrast improvement cannot be equated directly with anatomical fidelity or clinical utility.