Active learning using deep Bayesian networks for surgical workflow analysis.

Area of Science:

• Computer-assisted surgery
• Machine learning
• Surgical workflow analysis

Background:

• Machine learning, particularly convolutional neural networks (CNNs), is vital for surgical workflow analysis but requires extensive labeled data.
• Obtaining expert annotations for surgical data is a significant bottleneck, hindering the development of accurate models.

Purpose of the Study:

• To reduce the annotation effort required for training machine learning models in surgical workflow analysis.
• To adapt deep Bayesian networks (DBNs) for active learning in image- and video-based surgical workflow analysis.
• To compare different uncertainty metrics for effective data selection in active learning.

Main Methods:

• Implemented a deep Bayesian network (DBN) based active learning approach for surgical workflow analysis.
• Extended the DBN approach using a recurrent architecture for video-based analysis.
• Evaluated various uncertainty metrics, including entropy and variation ratio, to guide data selection.

Main Results:

• The DBN-based active learning approach significantly outperformed random data selection for surgical workflow analysis tasks.
• Metrics like entropy and variation ratio demonstrated consistent performance across instrument presence detection and surgical phase segmentation.
• The study successfully applied active learning to reduce the annotation burden for CNNs in surgical contexts.

Conclusions:

• Deep Bayesian network-based active learning strategies enable selective data annotation, significantly reducing the need for labeled training data.
• This approach is effective for various surgical workflow analysis tasks, enhancing the efficiency of machine learning model development.
• The findings pave the way for more accessible and efficient AI-driven tools in computer-assisted surgery.