Advancing cardiac MRI multi-structure segmentation: A semi-supervised multidimensional consistency constraint...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep convolutional neural networks (DCNNs) show promise for medical Magnetic Resonance Imaging (MRI) segmentation.
• Challenges remain in semantic discrimination, boundary delineation, and spatial context modeling for DCNNs in MRI segmentation.

Purpose of the Study:

• To introduce the Multidimensional Consistency Constraint Learning Network (MDCC-Net) for multi-structure cardiac MRI segmentation.
• To address limitations of existing DCNNs using a semi-supervised approach.

Main Methods:

• MDCC-Net utilizes a shared encoder and multiple differentiated decoders.
• The network incorporates pyramid boundary consistency features and spatial consistency constraints.
• Mutual consistency constraints, pseudo-labels, Dice loss, and mean squared error loss are employed to enhance segmentation.

Main Results:

• MDCC-Net achieved state-of-the-art performance on the ACDC cardiac MRI dataset for segmenting left ventricle (LV), myocardium (MYO), and right ventricle (RV).
• Average Dice coefficient reached 0.8763, Jaccard index 0.7906.
• Optimal Dice for LV was 0.8965 and Average Surface Distance for RV was 0.5391.
• Model generalization was confirmed on the M&Ms dataset.

Conclusions:

• MDCC-Net effectively enhances multi-structure cardiac MRI segmentation through multidimensional consistency constraints.
• This approach offers a foundation for integrating multifeature fusion in clinical automated and semiautomated segmentation.
• Potential improvements for diagnostic and treatment planning processes in clinical settings are highlighted.