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

Area of Science:

• Computer Vision
• Medical Training
• Surgical Skill Assessment

Background:

• Low-resource settings face a critical shortage of skilled surgeons.
• Computer-assisted surgical skill evaluation using video analysis is vital for alternative training.
• Existing tool detection models are complex and not optimized for in vitro training data.

Purpose of the Study:

• To evaluate various deep learning models for real-time surgical tool detection.
• To assess model performance on a custom laparoscopic box-trainer dataset.
• To identify models suitable for low-cost embedded devices in resource-constrained environments.

Main Methods:

• Tested multiple anchor-based and anchor-free, convolution- and transformer-based deep learning models.
• Evaluated models on an in-house laparoscopic box-trainer dataset.
• Focused on real-time performance metrics (inference time, FPS) on hardware like NVIDIA Jetson Orin Nano.

Main Results:

• YOLOv8-X demonstrated the highest accuracy with mAP50 of 99.5% and mAP50:95 of 96.6%.
• YOLOv8-X achieved an inference time of 23.5 ms (42.6 FPS) on the Jetson Orin Nano.
• YOLOv11-N was the most efficient, with an inference time of 3.1 ms (322.6 FPS), showing minimal accuracy drop.

Conclusions:

• Deep learning models can effectively perform real-time surgical tool detection for skill assessment.
• YOLOv8-X and YOLOv11-N show significant potential for surgical training in resource-constrained settings.
• The findings support the development of accessible, computer-assisted surgical training tools.