Mutual consistency learning for semi-supervised medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning models for medical image segmentation often struggle with limited annotations, leading to uncertainty in ambiguous regions.
• Leveraging challenging samples from unlabeled data can enhance the effectiveness of semi-supervised segmentation models.

Purpose of the Study:

• To propose a novel mutual consistency network (MC-Net+) for improved semi-supervised medical image segmentation.
• To effectively exploit unlabeled data by addressing model uncertainty in ambiguous regions.

Main Methods:

• MC-Net+ utilizes a shared encoder with multiple decoders employing different up-sampling strategies to estimate model uncertainty.
• A mutual consistency constraint is applied between decoder outputs and soft pseudo-labels to regularize training and improve robustness in challenging areas.

Main Results:

• MC-Net+ demonstrated superior performance compared to five state-of-the-art semi-supervised methods on three public medical datasets.
• Extension experiments confirmed the model's effectiveness under standard semi-supervised settings, establishing a new state-of-the-art.

Conclusions:

• The proposed MC-Net+ effectively leverages unlabeled data for semi-supervised medical image segmentation by addressing model uncertainty.
• The novel approach sets a new benchmark for semi-supervised medical image segmentation accuracy and robustness.