Style-Enhanced CycleGAN for Unpaired Image-to-Image Translation: A New Approach to Medical Image Fusion Refinement

Endoscopy remains the standard procedure for diagnosing a wide range of conditions within internal organs. Multispectral medical imaging has introduced new avenues for enhanced diagnostics. These techniques capture information beyond the visible spectrum, revealing subtle details of tissue and the microvascular system. In this study, we propose a novel framework for endoscopic image enhancement by combining useful information from both white light (WL) imaging and narrow band (NB) imaging techniques. A primary challenge in this process is the unavoidable misalignment of images captured under varied settings. To overcome this, we introduce an approach to generate precisely aligned input images for fusion. We employ a Style-enhanced CycleGAN framework for unpaired image-to-image translation to synthesize NB images from WL images, achieving improved detail preservation and style control. With corresponding WL and NB images established, we apply a multi-step image fusion where the input images are first decomposed into seven components across three iterations. Subsequently, an adaptive structure tensor and Gaussian filter-based pulse-coupled neural network (ST-G-PCNN) and maximum-based fusion strategies are applied to these components. Quantitative evaluation demonstrates improved translation performance (FID = 150.5, KID = 0.055) compared to baseline CycleGAN and recent state-of-the-art methods, while fused images exhibit enhanced perceptual quality, reflected by higher entropy and lower BRISQUE, NIQE, and PIQE scores. Subjective evaluation further indicates that fused images are consistently preferred by physicians for diagnostic assessment. Overall, the proposed framework offers a practical and effective solution for enhancing endoscopic visualization, with the potential to support more reliable clinical interpretation.