Variance Extrapolated Class-Imbalance-Aware Domain Adaptive Myocardial Segmentation in Multi-Sequence Cardiac MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Automated myocardial segmentation in cardiac MRI is crucial but challenging due to variations across vendors and protocols.
• Existing methods struggle with generalizing across different MRI sequences (cine, T1 mapping, LGE) and scanner types.

Purpose of the Study:

• To develop an unsupervised domain adaptation approach for robust myocardial segmentation across multi-vendor cardiac MRI data.
• To enable consistent segmentation performance across distinct MRI sequences without sequence-specific annotations.

Main Methods:

• Proposed an unsupervised domain adaptation framework utilizing a class-imbalance self-training strategy.
• Implemented a hardness-aware pseudo-labeling approach for iterative refinement of segmentation accuracy.
• Employed variance-guided vicinal feature extrapolation to mitigate data scarcity and enhance joint training.

Main Results:

• The proposed framework demonstrated superior performance compared to existing methods.
• Segmentation accuracy was significantly improved, as evidenced by Dice coefficient and Hausdorff distance metrics.
• The method enables cross-protocol cardiac evaluation without the need for sequence-specific manual annotations.

Conclusions:

• The unsupervised domain adaptation approach offers robust and generalizable myocardial segmentation for cardiac MRI.
• This technique addresses limitations of current methods, facilitating more efficient and accurate cardiac analysis across diverse datasets.
• The framework holds potential for improving automated diagnosis and treatment planning in cardiology.