Automatic Surgical Items Tracking in Laparoscopic Surgery Utilizing Deep Learning for Retained Items Detection and Classification

Abstract

During surgeries, some objects might be unintentionally retained in the human body. The most frequently retained item is the surgical gauze. All theatre rooms have adopted standard counting procedures to prevent surgical item retention. However, these procedures are prone to human errors and are, therefore, unreliable in some cases. In cases where items have been retained, the patients must be exposed to X-rays and unnecessary anesthesia, which is an additional risk to the patient. New technologies such as barcode readers and radiofrequency identification have been proposed to minimize human intervention in surgical item monitoring. However, these technologies still require human input and are sometimes complex for healthcare professionals. Advanced technologies such as deep learning could be used to automate surgical tools monitoring and tracking to minimize human errors. In this study, a performance comparison of various deep learning models in classifying images containing surgical tools is done. Four models are trained and validated using a binary class and multiclass datasets. The binary class dataset gives higher accuracy and lower loss than the multiclass dataset. YOLOv8 gives the highest accuracy among the tested models. The other models (MobileNet, ResNet and DenseNet) could be used for classification tasks in surgical applications depending on the complexity of the process.