Narrow-band imaging (NBI) improves visualization of mucosal microvasculature and is widely used for colorectal lesion characterization. However, variations in vascular contrast and structural clarity can limit both clinical interpretation and automated analysis. In this work, we propose a customized CycleGAN framework to enhance NBI images using generators trained through unpaired translation between white light (WL) and NBI modalities. During training, the generators learn modality-specific mappings that capture vascular and mucosal characteristics of NBI. We hypothesize that applying this learned transformation to real NBI images can reinforce subtle vascular patterns and improve structural representation. To better preserve fine anatomical details, the CycleGAN architecture was modified by replacing mean absolute error with structural similarity index (SSIM) in cyclic consistency loss and introducing feature concatenation within generator residual blocks. Two generator configurations with 6 and 9 residual blocks were evaluated. The enhanced images showed increased entropy and chromatic contrast relative to the original NBI images. The deeper model demonstrated improved segmentation performance across diverse segmentation architectures, achieving higher Dice coefficients and Jaccard indices with comparable F2 scores. Further lesion class-wise results indicate that it can effectively enhance NBI images and consistently improve lesion segmentation across lesion classes and thereby classification performance.
Colonoscopy can achieve significantly enhanced polyp detection by employing narrowband imaging (NBI) along with conventional white light (WL) imaging. However, problems associated with switching between two modalities during colonoscopy hinder its widespread clinical application. This study proposes a novel skip-connected residual layer-based customized CycleGAN model, which we refer to as customized CycleGAN with balanced skip-networked residual layer (CCGAN-BSNRL), for achieving an effective translation of WL colonoscopy images into NBI-like images. We chose the residual layer for customization as it is crucial for capturing the core spatial features of the input and translating them into the target domain. To evaluate, we have chosen a set of existing metrics to compare images generated by our model with images generated by other related models. We additionally propose those metrics for the comprehensive evaluation of other generative models’ generated images. Finally, a simple segmentation model is used to inspect the improvement in polyp detection. Considering the overall results, with a significantly enhanced detection of polyps achieved through the translated NBI (Dice score greater than 6%, Jaccard score greater than 8%, on average) from the proposed model, it is expected that the CCGAN-BSNRL model should positively impact the diagnosis accuracy of colorectal cancer in clinical applications.