DS-SurgNet:Dense Scale-aware Surgical Network

Surgical instrument segmentation is critical for computer-assisted surgery systems, directly affecting intraoperative guidance accuracy. However, challenges such as specular reflections, variable illumination, and indistinct boundaries limit the performance of existing methods. We propose DS-SurgNet, a novel illumination-adaptive framework that makes five key contributions: (1) adaptive illumination correction for robust lighting normalization, (2) illumination-adaptive boundary enhancement with specialized edge detection, (3) deformable convolutional receptive field blocks for handling irregular instrument morphologies, (4) illumination-texture integrated attention mechanism for enhanced feature discrimination, and (5) multi-scale adaptive feature aggregation for comprehensive representation learning. Our framework uniquely integrates these innovations through a boundary-aware loss function using Laplacian operators that specifically penalizes segmentation boundary errors. Comprehensive evaluations demonstrate that DS-SurgNet achieves 81.8% IoU and 89.1% Dice score, representing significant improvements over other methods, particularly under challenging illumination conditions and complex instrument configurations, thereby enabling more reliable real-time surgical guidance systems.