A discontinuity-preserving regularization for deep learning-based cardiac image registration.

Area of Science:

• Medical imaging
• Computational anatomy
• Deep learning

Background:

• Sliding motion between organs due to respiratory and cardiac activity is often overlooked in image registration.
• This neglect leads to suboptimal registration accuracy in medical imaging.
• Accurate registration is crucial for various clinical applications, including diagnosis and treatment planning.

Purpose of the Study:

• To develop novel methods for handling discontinuities at organ boundaries in medical image registration.
• To improve the accuracy and robustness of deep learning-based image registration algorithms.
• To enhance the performance of cardiac image registration by addressing motion-induced artifacts.

Main Methods:

• Introduced a discontinuity-preserving regularization (DPR) term that maintains local discontinuities by relaxing constraints at organ boundaries identified via segmentation masks.
• Proposed a weakly supervised mask dissimilarity loss (MDL) to measure similarity between fixed and deformed image masks.
• Integrated DPR and MDL into the loss function for network training, with no need for masks during inference.

Main Results:

• The DPR term improved Dice coefficients by up to 0.081 across different registration networks (CRNet, VoxelMorph, ViT-V-Net).
• Significant enhancements were observed in Hausdorff Distance and Average Surface Distance metrics.
• MDL provided a slight improvement (within 1%) in quantitative indicators, leading to more reasonable displacement fields and improved image fidelity.

Conclusions:

• The proposed DPR and MDL regularization terms effectively address organ boundary discontinuities in medical image registration.
• These methods significantly enhance the accuracy of deep learning-based cardiac image registration.
• The regularization terms are generic and can be extended to various registration networks.