Medical image segmentation using dual-decoder mutual teaching with a mean teacher framework

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Accurate medical image segmentation is crucial for clinical applications.
• Manual pixel-level annotation of medical images is time-consuming and labor-intensive.
• Deep learning models require large annotated datasets for optimal performance.

Purpose of the Study:

• To develop a novel semi-supervised segmentation method to reduce manual annotation effort.
• To improve the stability and shape consistency of deep learning models in segmentation tasks.
• To achieve promising segmentation performance with limited labeled and abundant unlabeled images.

Main Methods:

• Introduced dual-decoder mutual teaching (DDMT), a semi-supervised segmentation method.
• Incorporated smoothed exponential moving average (sEMA) for enhanced model stability.
• Integrated shape consistency constraint (SCC) for consistent shape learning across decoders.

Main Results:

• DDMT demonstrated promising segmentation performance on limited labeled data.
• The method consistently outperformed state-of-the-art semi-supervised learning methods.
• Experiments on left atrium, pancreas, and optic disc datasets validated DDMT's effectiveness.

Conclusions:

• DDMT offers an effective solution for reducing manual annotation in medical image segmentation.
• The proposed method enhances model stability and shape consistency.
• DDMT shows significant potential for clinical applications requiring accurate image segmentation.