Towards Automated Performance Assessment for Laparoscopic Box Trainer using Cross-Stage Partial Network

Recent advances in laparoscopic surgery have increased the need to improve surgical resident training and feedback by incorporating simulator based training in traditional training programs. However, the current training methods still require the presence of an expert surgeon to assess the surgical dexterity of the trainee. This process is time consuming and may lead to subjective assessment. This research aims to extend the application of object detection in laparoscopy training by tracking tool motion, surgical object detection and tracking. YOLOv5 and scaled-YOLOv4 object detection neural networks based on cross-stage partial network (CSP) are trained and tested on the Fundamentals of Laparoscopic Surgery (FLS) pattern cutting exercise in a box trainer. Experiments show that Scaled-YOLOv4 have a mAP score of 98.9, 79.5 precision and 98.9 recall for bounding boxes on a limited training dataset. This research clearly demonstrates the potential of using CSP networks in automated tool motion analysis for the assessment of the resident's performance during training.