Real-Time Tool Detection in Laparoscopic Datasets for Surgical Training in Low-Resource Settings

In low-resource settings, there is a critical need for skilled surgeons. Alternative training processes that include computer-assisted surgical skill evaluation are essential to address this gap. Using tool detection, surgical videos can be leveraged to derive insights into surgical skill assessment. However, state-of-the-art laparoscopic tool detection methods usually have more complex architectures tailored for in vivo data, which suffer from challenges such as smoke, occlusion, bleeding, etc., which are absent from in vitro training contexts. Thus, this paper tests multiple anchor-based and anchor-free, convolution- and transformer-based, traditional (non-surgical domain-specific) computer vision deep learning state-of-the-art models. With various hardware configurations on a newly curated in-house laparoscopic box-trainer dataset, we emphasise real-time performance on low-cost embedded devices. Overall, the anchor-free YOLOv8-X model was the most accurate, achieving ${\mathrm{mAP}}_{50}$ of 99.5% and ${\mathrm{mAP}}_{50:95}$ of 96.6% with an inference time of 23.5 ms/ $\approx$ 42.6 FPS on an NVIDIA Jetson Orin Nano 8GB (comparable low-cost hardware which could be expected to run real-time skill assessment methods for surgical training boot camps in a resource-constrained environment). The most efficient model was YOLOv11-N, providing 3.1 ms/ $\approx$ 322.6 FPS with a performance difference of +0% ${\mathrm{mAP}}_{50}$ and -2.1% ${\mathrm{mAP}}_{50:95}$ . The results highlight the models' potential for effective real-time detection of surgical tools and are suitable for further downstream assessment of surgical skills, even in resource-constrained environments.