Data-centric multi-task surgical phase estimation with sparse scene segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Surgical Workflow Analysis

Background:

• Surgical workflow estimation segments videos into actions or phases for applications like real-time feedback and automated reporting.
• Traditional methods often decouple feature extraction and temporal fusion due to computational constraints.
• Existing research primarily focuses on temporal model development for phase estimation.

Purpose of the Study:

• To develop a data-centric training pipeline for surgical phase estimation.
• To maximize the utility of existing, often isolated, surgical video datasets.
• To propose a multi-task learning approach for accurate surgical phase prediction.

Main Methods:

• Utilized dense phase annotations from Cholec80 dataset.
• Incorporated sparse scene segmentation annotations (instrument and anatomy) from CholecSeg8k.
• Developed a multi-task encoder to fuse both annotation streams based on importance.
• Jointly optimized the encoder for accurate phase prediction.

Main Results:

• Achieved comparable results to state-of-the-art complex architectures with a relatively simple model.
• Demonstrated the effectiveness of using a small fraction of scene segmentation annotations.
• Validated the performance in similar evaluation settings as previous works.

Conclusions:

• A data-centric approach with a simple multi-task model can yield high performance in surgical phase estimation.
• Effective fusion of diverse annotation types enhances model accuracy.
• This methodology encourages research prioritizing data utilization alongside model development.