A comparative evaluation of spatial pooling methods for surgical tool detection

Abstract

Surgical tool detection is an important aspect for recognising surgical activities and understanding surgical workflow. Laparoscopic videos represent an information source that can be used for recognising surgical tools. However, manual labelling of tool incidence and location in such data is extremely time intensive. Therefore, weaklysupervised approaches have been developed to perform tool localisation. In this study, three types of spatial pooling methods were implemented to evaluate the influence of each method on the performance of weakly-supervised model. The best achieved performance was a mean average precision (mAP) of 94% for tool classification and a f1-score of 70% for tool localisation. Experimental results showed the importance of selecting an appropriate pooling function to enhance model performance.

Keywords

• convolutional neural network (cnn)
• surgical tool localisation
• laparoscopic videos
• weakly-supervised learning