Enhancing endoscopic detection of gastrointestinal lesions: virtual narrow band imaging for improved CAD systems (Conference Presentation)

White light endoscopy (WLE) is a widely utilized technique for diagnosing and detecting gastrointestinal lesions. Narrow band imaging (NBI) enhances the visibility of vascular patterns and mucosal structures. However, the implementation of NBI in some hospitals is limited by resource constraints, leading to a reliance on WLI. To leverage the advantage NBI cost-effectively and accessibly, this study proposes an image-to-image translation from WLI to virtual NBI (VNBI) images utilizing unpaired cycle-consistent Generative Adversarial Networks (CycleGANs). By incorporating this generator as a preprocessing step, we developed detection models to assess the efficacy of VNBI translation for lesion classification. Comparative analysis of the models, with and without VNBI image augmentation, indicated that the model trained on VNBI images achieved superior performance in lesion classification. Our findings suggest that VNBI images enhance the performance of WL datasets, improving the overall accuracy of CAD systems in endoscopic imaging.