Triple-task mutual consistency for semi-supervised 3D medical image segmentation.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Semi-supervised deep learning is crucial for medical image segmentation.
• Multi-task learning with consistency regularization shows promise but often underutilizes Signed Distance Maps (SDM) for edge awareness.
• Existing methods can create over-dependence between segmentation tasks.

Purpose of the Study:

• To propose a novel Triple-Task Mutual Consistency (TTMC) framework.
• To enhance shape and edge awareness in medical image segmentation.
• To overcome task dependence issues in multi-task learning for segmentation.

Main Methods:

• Developed a Triple-Task Mutual Consistency (TTMC) framework.
• Introduced Signed Attention Maps (SAM) as an auxiliary task to improve edge awareness.
• Implemented a deep network for joint prediction of classification maps, SDM, and SAM.
• Utilized an optimized differentiable transformation layer and task-level consistency regularization with unlabeled data.

Main Results:

• The TTMC framework demonstrated significant performance gains on public Left Atrium and NIH Pancreas datasets.
• Achieved superior results compared to state-of-the-art semi-supervised segmentation methods.
• Effectively leveraged unlabeled data to enhance segmentation accuracy and edge awareness.

Conclusions:

• The proposed TTMC framework effectively enhances shape and edge awareness in medical image segmentation.
• This approach overcomes limitations of previous methods by better utilizing SDM and reducing task dependence.
• TTMC offers a promising direction for advancing semi-supervised medical image segmentation techniques.