Surgical Tool Classification & Localisation Using Attention and Multi-feature Fusion Deep Learning Approach
Year: 2023
Type: article
Abstract: Analysing laparoscopic videos, particularly for surgical tool classification and localisation, has attained interest in the field of surgical data science since they represent an extensive information source. However, the difficulties for acquiring and labelling these videos have led to paucity of labelled datasets. Consequently, the progress of developing robust and generalised surgical tool detection models was slowed down, and translating these models into the medical field was hindered. In this work, supervised surgical tool classification and weakly-supervised tool localisation in laparoscopic videos were addressed. A base convolutional neural network (CNN) model was adapted to perform both tasks by incorporating multi-map localisation layers. Squeeze-and-excitation modules were added to the CNN to enhance the ability of the model to generate better focused informative features. Additionally, features at multiple stages of the CNN were combined and fused in a batch normalisation layer to enhance model generalisability. The proposed model was evaluated on the popular Cholec80 dataset. Experimental results of 94.1% mean average precision for tool classification and 70.1% F1-score for tool localisation revealed the ability of the model to learn better features for both tasks. The proposed approach showed the advantages of integrating attentions modules and multi-stage features fusion technique for surgical tool classification and localisation.
Source: IFAC-PapersOnLine
Language: English
Subfield: Surgery
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W4388917953
