All-Around Real Label Supervision: Cyclic Prototype Consistency Learning for Semi-Supervised Medical Image...

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Semi-supervised learning significantly advances medical image segmentation by reducing annotation costs.
• Consistency-based methods are prominent, using real labels for supervised loss and unlabeled data for unsupervised consistency.
• Existing methods do not fully exploit the rich supervision signals within expert-examined real labels for unlabeled data.

Purpose of the Study:

• To investigate exploiting unlabeled data through explicit real label supervision in semi-supervised training.
• To develop a novel framework that enhances medical image segmentation by maximizing the utility of available annotations.
• To introduce a new paradigm for consistency learning that utilizes all available label information.

Main Methods:

• Proposing a novel Cyclic Prototype Consistency Learning (CPCL) framework based on prototypical networks.
• Implementing a labeled-to-unlabeled (L2U) prototypical forward process and an unlabeled-to-labeled (U2L) backward process.
• Transforming 'unsupervised' consistency into 'supervised' consistency by enabling explicit real label supervision for all data.

Main Results:

• The CPCL framework encourages more discriminative and compact features, enhancing the segmentation network.
• Experiments on brain tumor and kidney segmentation demonstrate effective exploitation of unlabeled data.
• CPCL outperforms existing state-of-the-art semi-supervised medical image segmentation methods.

Conclusions:

• The proposed CPCL framework achieves 'all-around real label supervision' by effectively utilizing unlabeled data.
• This approach offers a significant improvement over traditional consistency-based methods in medical image segmentation.
• CPCL shows strong potential for advancing semi-supervised learning in medical image analysis.