Lightweight YOLOv8 variants for enhanced real-time surgical tool detection

Accurate and real-time surgical tool detection is essential in minimally invasive surgery (MIS) to ensure safe and effective computer-assisted interventions (CAI) and robot-assisted surgery (RAS). However, achieving high accuracy under the strict computational constraints of embedded surgical platforms remains a significant challenge. This article introduces a set of enhanced lightweight You Only Look Once (YOLO)v8 variants specifically tailored for real-time surgical tool localization. The proposed architectures integrate several key innovations: Ghost Convolution for efficient feature extraction, a C2f-Ghost module for compact representation, and the SC3T module, which synergizes Transformer blocks with Spatial Pyramid Pooling. Additionally, attention mechanisms—including the Context Augmentation Module (CAM) and Convolutional Block Attention Module (CBAM)—are incorporated to refine feature discriminability. Furthermore, the Scylla-Intersection over Union (SIoU) loss is employed to optimize bounding box regression for elongated instruments. Evaluations on the public m2cai16-tool-locations dataset demonstrate that the best variant achieves 95.7% mAP@0.5 while reducing parameters and Giga FLOating-point Operation(s) Per Seconds (GFLOPs) by up to 3× and 1.8×, respectively, compared to the YOLOv8 baseline. These results indicate that our design offers a competitive accuracy-efficiency trade-off, facilitating practical deployment in resource-constrained, real-time surgical systems.