Exploring Anchor-Free Object Detection for Surgical Tool Detection in Laparoscopic Videos: A Comparative Study of CenterNet++ and Anchor-Based Models

Abstract

Minimally Invasive Surgery (MIS) has transformed modern medicine, offering reduced re covery times, minimal scarring, and lower risks of infection. However, MIS procedures alsopresent unique challenges, particularly in visualizing and manipulating surgical tools within a limited field of view. As a solution, this thesis investigates anchor-free deep learning mod els for real-time surgical tool detection in laparoscopic videos, proposing CenterNet++ as apotential improvement over traditional anchor-based methods. The hypothesis guiding thiswork is that anchor-free detectors, by avoiding predefined anchor boxes, can more effectively handle the diverse shapes, sizes, and positions of surgical tools. The primary objective of this thesis is to evaluate the performance of CenterNet++ in surgical tool detection compared to popular anchor-based models, specifically Faster R-CNN and YOLOv4, using the m2cai16-tool-locations dataset. CenterNet++ is examined in dif ferent configurations—including complete and real-time optimized (Fast-CenterNet++) ver sions—and tested against Faster R-CNN and YOLOv4 to assess trade-offs in accuracy and efficiency. Experimental results demonstrate that while CenterNet++ achieves high precision, particularly in scenarios requiring meticulous localization, its inference speed is significantly slower than YOLOv4, which attained real-time speeds at 128 FPS. CenterNet++’s unique keypoint refinement mechanism, though beneficial for localization, impacts its computational efficiency, highlighting areas for further optimization. To bridge this gap, several architectural improvements are proposed based on YOLOv4’s streamlined design. These include integrating modules like Spatial Pyramid Pooling (SPP) and Path Aggregation Network (PANet), along with reducing input resolution in the Fast CenterNet++ configuration. Additionally, future work is suggested to explore CenterNet++ in larger, more complex datasets and to develop semi-supervised learning approaches that could mitigate the limitations of annotated surgical datasets. In conclusion, this thesis contributes a comprehensive evaluation of anchor-free models for surgical tool detection, providing a foundation for further advancements in real-time, high precision object detection for surgical assistance. The findings underscore the potential of anchor-free models, such as CenterNet++, to meet the evolving demands of MIS with targeted architectural adaptations.