Min-Max Similarity: A Contrastive Semi-Supervised Deep Learning Network for Surgical Tools Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Limited pixel-level annotated data hinders neural network training for medical image segmentation.
• Existing semi-supervised methods struggle with accuracy and generalizability.

Purpose of the Study:

• To develop a semi-supervised segmentation network that effectively utilizes unlabeled data.
• To improve the accuracy and efficiency of surgical tool segmentation in medical imaging.

Main Methods:

• Proposed a semi-supervised segmentation network leveraging contrastive learning.
• Introduced Min-Max Similarity (MMS) for dual-view training with classifiers and projectors.
• Utilized all-negative and positive/negative feature pairs for supervised learning and consistency measurement.

Main Results:

• The proposed method outperformed state-of-the-art semi-supervised and fully supervised segmentation algorithms on multiple datasets.
• The algorithm successfully recognized unknown surgical tools and provided accurate predictions.
• Achieved real-time inference speeds of approximately 40 frames per second.

Conclusions:

• The MMS-based semi-supervised network is a viable solution for medical image segmentation with limited annotations.
• The approach demonstrates strong performance in recognizing diverse surgical tools and supports real-time applications.
• This method offers a significant advancement in automated surgical tool segmentation.