SegMatch: semi-supervised surgical instrument segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Assisted Surgery
• Machine Learning

Background:

• Surgical instrument segmentation is crucial for advanced surgical assistance and computer-assisted interventions.
• Accurate segmentation requires large annotated datasets, which are expensive and time-consuming to create.

Purpose of the Study:

• To introduce SegMatch, a semi-supervised learning method for surgical instrument segmentation.
• To reduce the reliance on extensive manual annotation for laparoscopic and robotic surgical images.

Main Methods:

• SegMatch adapts the FixMatch semi-supervised learning pipeline for segmentation tasks.
• It utilizes consistency regularization and pseudo-labelling with weakly and strongly augmented images.
• A trainable adversarial augmentation strategy is incorporated to enhance augmentation relevance.

Main Results:

• SegMatch outperforms fully-supervised methods by effectively leveraging unlabelled data.
• The method surpasses existing state-of-the-art semi-supervised models across various data ratios.
• Evaluated on MICCAI Instrument Segmentation Challenge, Robust-MIS 2019, EndoVis 2017, and CholecInstanceSeg datasets.

Conclusions:

• SegMatch offers a viable solution to reduce annotation burden in surgical image segmentation.
• The proposed method demonstrates significant improvements in segmentation accuracy with limited labelled data.
• This advancement holds promise for more accessible and effective computer-assisted surgical systems.