Ambiguity-selective consistency regularization for mean-teacher semi-supervised medical image segmentation.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning significantly advances medical image segmentation by reducing reliance on expert annotations.
• The mean-teacher model, a perturbation consistency learning approach, is a common baseline for this task.
• Existing methods often overlook the crucial aspect of selecting informative consistency targets from unlabeled data.

Purpose of the Study:

• To improve the mean-teacher model by introducing a novel ambiguity-consensus mean-teacher (AC-MT) model.
• To leverage informative complementary clues from ambiguous regions within unlabeled medical image data.
• To enhance the efficiency and accuracy of semi-supervised medical image segmentation.

Main Methods:

• Developed and benchmarked a family of plug-and-play strategies for ambiguous target selection based on entropy, model uncertainty, and label noise self-identification.
• Incorporated an estimated ambiguity map into the consistency loss function to promote consensus in informative regions.
• Focused on learning from the perturbed stability of these identified informative regions within unlabeled data.

Main Results:

• The AC-MT model demonstrated substantial improvements over current state-of-the-art methods in left atrium and brain tumor segmentation.
• Ablation studies validated the proposed strategies and showcased impressive performance under various extreme annotation conditions.
• The method effectively identifies and utilizes the most valuable voxel-wise targets from unlabeled data.

Conclusions:

• The proposed ambiguity-consensus approach significantly enhances semi-supervised medical image segmentation.
• Focusing on ambiguous regions and perturbed stability offers a more effective learning strategy.
• AC-MT provides a robust and adaptable framework for medical image segmentation with limited annotations.