Voxel-wise adversarial semi-supervised learning for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning reduces annotation costs in medical imaging.
• Existing methods use consistency regularization, pseudo-labeling, or adversarial learning.
• Previous approaches often focus on local features or single-class relationships.

Purpose of the Study:

• To develop a novel adversarial learning-based semi-supervised segmentation method.
• To effectively embed local and global features from multiple layers.
• To learn context relations across multiple classes for improved medical image segmentation.

Main Methods:

• Introduced a novel adversarial learning framework for semi-supervised segmentation.
• Utilized a voxel-wise feature discriminator considering multi-layer voxel-wise features.
• Improved representation learning to overcome information loss and enhance stability.

Main Results:

• Demonstrated effectiveness on single-class and multi-class segmentation tasks using benchmark datasets.
• Achieved superior performance compared to state-of-the-art semi-supervised learning approaches.
• Improved network performance by 2% in Dice score coefficient on a multi-organ dataset by leveraging unlabeled data.

Conclusions:

• The proposed method effectively embeds class-specific features across diverse medical datasets.
• Visualizations confirm a well-distributed and separated feature space, enhancing prediction accuracy.
• The adversarial learning approach successfully leverages unlabeled data for improved medical image segmentation.