Neural Network Classification of Surgical Tools in Gynecological Videos

Abstract Automated surgical tool classification will improve the workflow of surgery. Previous research tackled this task mainly in cholecystectomy procedures due to availability of a relatively large and labelled set (Cholec80 dataset). However, the complexity of the procedure type has an impact on the robustness of the deep learning approaches. Therefore, the classification capability of CNNs on data of more complex procedures with many surgical tools was investigated. In this work, laparoscopic videos of 14 gynaecological procedures were recorded and labelled for surgical tool presence. Then, the DenseNet-121 model was trained to identify surgical tools according to functionality. Experimental results imply high classification performance for some surgical tools. The mean average precision over all the tools was 67%. This study is an initial benchmark for detecting surgical tools in realistic settings.