HAMSRNet-Hybrid Attention Multiscale Super-Resolution Network for Endoscopic Images

Abstract

Stereo endoscopy is now a key part of minimally invasive surgery, giving surgeons clearer depth cues and better awareness of surrounding anatomy. Still, the cameras used in these systems often produce low-resolution images with noise and blur, which can hide small but important structures such as thin vessels, tissue boundaries, and surgical tools. These limitations affect the accuracy and confidence needed during delicate procedures. Many existing stereo super-resolution (SR) approaches improve visual detail, but most are too computationally demanding for real-time use in the operating room.To address these constraints, we present HAMSRNet, a lightweight hybrid network designed to enhance stereo endoscopic images while keeping computation low. The model is built around three main components. The first is a Lite Feature Extractor that relies on Ghost and depthwise separable convolutions together with a simple squeeze gate to collect rich features efficiently. The second is the XView-LiteFusion module, which aligns left and right views through parallax cues and shared residual blocks while using adaptive attention to maintain stereo consistency. The third is an Edge-Aware Reconstruction Head that gathers multiscale information and applies channel–spatial attention and efficient upsampling to recover fine anatomical structure. Experiments on the da Vinci, SCARED, and Kidney Boundary datasets show that HAMSRNet offers higher PSNR and SSIM than both lightweight and full-scale SR models. Visual results further highlight improvements in vessel continuity, edge definition, and overall structural realism. Some performance drop is observed under extreme occlusion or challenging lighting, and future work will explore uncertainty modelling and self-supervised refinement to handle these cases.Clinical relevance—HAMSRNet provides sharper and more reliable endoscopic views while keeping computation low enough for real-time deployment in robotic and laparoscopic systems. By improving the visibility of subtle boundaries and textures, the method supports safer and more precise navigation and decision-making during surgery.