Exploring Anchor-Free Object Detection Models for Surgical Tool Detection: A Comparative Study of Faster-RCNN, YOLOv4, and CenterNet++

The evolution of surgical techniques has led to the widespread adoption of Minimally Invasive Surgeries (MIS), offering benefits such as reduced scarring, shorter hospitalization, and less post-operative pain. Despite these advantages, laparoscopic procedures present challenges including limited vision and maneuverability of surgical tools, requiring advanced assistive technologies. This paper presents a comparative study of three state-of-the-art object detection models—Faster-RCNN, YOLOv4, and CenterNet++—for surgical tool detection in laparoscopic videos, using the m2cai16-tool-locations dataset. Our aim is to explore the potential of anchor-free models, exemplified by CenterNet++, and compare them against widely used anchor-based models. The results demonstrate that CenterNet++ achieves competitive performance, with a mAP50 of 0.926, mAP75 of 0.587, and mAP50:95 of 0.556, particularly excelling in higher IoU thresholds and specific surgical tool classes. YOLOv4 demonstrated the best performance in lower IoU thresholds, achieving the highest mAP50 of 0.950. These findings highlight the potential of anchor-free models in the medical imaging domain and suggest promising avenues for further research and application in surgical tool detection.