MDT: semi-supervised medical image segmentation with mixup-decoupling training.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Semi-supervised segmentation is crucial in medicine but hindered by scarce expert-annotated data.
• Existing methods often inconsistently process labeled/unlabeled data, discarding valuable information and ignoring pseudo-labeling biases.
• Challenges include limited data perturbations, confirmation bias, and class imbalance in pseudo-labeling.

Purpose of the Study:

• To propose a novel semi-supervised medical image segmentation method, Mixup-Decoupling Training (MDT).
• To address the limitations of existing methods in handling scarce annotated data and pseudo-labeling issues.
• To improve segmentation performance and reduce the demand for manual data annotation.

Main Methods:

• Introduced 'mixup-decoupling' perturbation strategy for data regularization at both data and feature levels.
• Established a dual learning paradigm combining consistency and pseudo-labeling.
• Employed categorical entropy filtering for high-confidence pseudo-label selection.

Main Results:

• MDT demonstrated competitive segmentation performance on 2D and 3D datasets.
• Experimental results showed MDT outperforms state-of-the-art semi-supervised segmentation methods.
• Achieved superior quantitative and qualitative results compared to existing approaches.

Conclusions:

• MDT significantly reduces the need for manually labeled medical data, easing annotation difficulties.
• The method offers a new, adaptable approach for semi-supervised learning in medical imaging.
• MDT advances research in semi-supervised learning and computer-aided diagnosis technologies.