Medical image segmentation with generative adversarial semi-supervised network.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation requires extensive annotated data, which is scarce and costly to obtain.
• Limited annotations pose a significant challenge for achieving high performance in segmentation tasks.

Purpose of the Study:

• To develop a self-learning method for medical image segmentation using limited annotations.
• To improve segmentation accuracy by effectively utilizing unlabeled medical images.

Main Methods:

• Proposing a generative adversarial semi-supervised network (GASSN).
• Utilizing limited annotated images for primary supervision.
• Employing unlabeled images with generated pseudo-labels for auxiliary training.
• Training an uncertainty discriminator to assess pseudo-label reliability.
• Applying feature mapping loss for statistical distribution consistency.

Main Results:

• The GASSN method achieved high Dice coefficients across multiple datasets (e.g., 0.8402-0.9121 on the right ventricle dataset).
• Performance improvements reached up to 28.6 points higher than fully supervised baselines.
• Demonstrated effectiveness with varying proportions of annotated data (1/8 to 1/2).

Conclusions:

• The proposed generative adversarial semi-supervised network effectively addresses the challenge of limited annotations in medical image segmentation.
• This self-learning approach significantly enhances segmentation performance by leveraging unlabeled data.
• The method shows broad applicability across diverse medical imaging datasets.