Dual cross-attentive mutual teaching for semi-supervised 3D medical segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Large labeled datasets are crucial for 3D medical image segmentation.
• Semi-supervised learning offers a solution to reduce reliance on extensive labeled data.
• Existing methods may not fully leverage unlabeled data for improved segmentation performance.

Purpose of the Study:

• To propose a novel Dual Crossed Attention Mutual Teaching (DCA-MT) framework for semi-supervised 3D medical image segmentation.
• To effectively integrate labeled and unlabeled data through advanced feature alignment and knowledge distillation.
• To enhance segmentation accuracy and robustness by enabling collaborative learning between teacher and student networks.

Main Methods:

• A two-branch VNet architecture with a co-evolving teacher-student network.
• High-dimensional feature alignment using Maximum Mean Difference (MMD) loss and contrast constraints.
• Semantic-level crossed attention and bidirectional knowledge distillation for inter-network feature exchange.

Main Results:

• The DCA-MT framework demonstrated superior performance on left atrial and pancreatic datasets.
• Experiments validated the effectiveness of high-dimensional feature alignment and cross-attention mechanisms.
• The bidirectional knowledge distillation significantly improved segmentation outcomes.

Conclusions:

• The proposed DCA-MT framework effectively utilizes both labeled and unlabeled data in 3D medical image segmentation.
• The integration of feature alignment, cross-attention, and mutual distillation enhances segmentation accuracy and robustness.
• DCA-MT offers a promising approach for semi-supervised learning in medical imaging applications.