Analysing attention convolutional neural network for surgical tool localisation: a feasibility study

Abstract Image-based surgical tool localisation and detection play an important role in developing intelligent systems for operating theatres of the future. In literature, proposed approaches require large amounts of data that are fully annotated with tool positions in the image. In this paper, a deep learning framework, trained on binary tool presence, was evaluated for surgical tool localisation in laparoscopic images. Gradient class activation maps (Grad-CAMs) were extracted from an attention-CNN model. The Grad-CAMs were then processed to generate bounding boxes over the surgical tools. Experimental results showed better performance of the attention-CNN compared to the base CNN model with mean tool localisation precision of 72.4% and 28.3%, respectively. These results show the potential of using attention modules to improve tool localisation in laparoscopic images.