A Minimax Optimization Framework for Attention Enhanced Surgical Instrument Detection in Laparoscopic Surgery

Abstract

Reliable detection of surgical instruments in endoscopic video is a foundational requirement for computer assisted intervention systems, supporting downstream tasks such as workflow recognition, skill assessment, and intra-operative safety monitoring. Endoscopic scenes are visually demanding: instruments are repeatedly occluded by tissue and blood, deform under articulation, exhibit specular highlights from metallic surfaces, and operate against backgrounds whose texture closely resembles instrument shafts. Conventional attention designs handle the spatial and channel axes either in isolation or in a fixed sequence, which leaves cross-axis interactions only weakly modelled and limits discrimination when instrument cues are subtle. This paper introduces GAR-DETR, an end-to-end detector specialised for laparoscopic abdominal surgery that augments an RT-DETR backbone with a single Gated Attention Refinement (GAR) module. The module evaluates two parallel attention pathways over the encoder feature map — one along spatial positions and one along feature channels — and fuses them through learnt sigmoid gates so that the more discriminative axis is emphasised at every location. A short pseudocode procedure summarises the operation. On the public M2CAI2016-Tool Locations benchmark, GAR-DETR achieves 98.83% mAP@0.5 across seven instrument classes, outperforming twenty competing detectors that include recent YOLO variants and transformer-based baselines, with the largest per-class improvements reported on Hook, Scissor, and Clipper.