Abstract
A retained foreign body is a medical error wherein surgical items are inadvertently left in the patient after surgery. In these cases, the patient requires reoperation with high risks of unprecedented death which incurs substantially high medico-legal costs. One of the risk factors observed to be associated with retained foreign bodies is the incorrect count of surgical instruments or sponges used. Modern technologies that have been developed to assist in surgical counts uses handheld scanners and radio frequency identification tags or barcodes embedded on the surgical item. This study proposes a vision-based approach to eliminate the use of handheld scanners and embedded tags on surgical items by employing computer vision with machine learning. In this new approach, the Single Shot Multibox Detector (SSD) with MobileNet is trained to detect common surgical items. The training was done in three iterations, while expanding the dataset on each iteration. Model 1 achieved the highest mean average precision of 87.12% upon evaluation on the test dataset. A surgical counter application was implemented based on the trained object detector and was evaluated on a series of videos. In terms of classification accuracy, model 2 performed better with an accuracy of 49.50%. In implementing the surgical counter application, the hybrid approach of combining object detection and object tracking significantly improved the speed by up to 200% with the fastest fps at 38.39 achieved by model 2.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have