Segmentation-Guided Colorization of Single-Channel Laser Illumination Images for Improved Surgical Navigability
Abstract
Laser speckle contrast imaging (LSCI) enables real-time intraoperative perfusion assessment in laparoscopic surgery, but visible wavelength LSCI systems require switching to laser illumination mode, producing a single-channel image that strips away the chromatic tissue cues surgeons rely on for anatomical navigation. Existing unpaired image translation methods lack semantic awareness, producing colorizations that are visually inconsistent across tissue types. We propose a segmentation-guided unpaired image-to-image translation framework that colorizes single-channel red-illumination laparoscopic images into perceptually natural RGB views, restoring visual navigability without interfering with the perfusion computation pipeline. We inject anatomical masks at every decoder layer for spatially consistent synthesis and introduce a class-wise contrastive loss that restricts patch sampling to semantically homogeneous regions, yielding harder within-class negatives. The discriminator is additionally conditioned on the mask for class-aware adversarial feedback, supplemented by a Gram matrix style loss. Our method is evaluated on the Dresden Surgical Anatomy Dataset in a fully unpaired setting, with the red channel serving as the source domain. Quantitatively, it achieves SSIM of 0.81, MS-SSIM of 0.87, PSNR of 22.51 dB, and LPIPS of 0.14, outperforming all baselines across all evaluated metrics. This work demonstrates improved colorization fidelity for laser-mode laparoscopic views under a simulated red-channel proxy, a step toward supporting surgical decision-making during perfusion-guided procedures.
