Anatomy-Regularized Representation Learning for Cross-Modality Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Unsupervised cross-modality synthesis is used to address limited labels in medical image segmentation.
• Current methods often fail to preserve high-level semantic information due to reliance on voxel-wise consistency.

Purpose of the Study:

• To propose a novel anatomy-regularized representation learning approach for segmentation-oriented cross-modality image synthesis.
• To improve the preservation of anatomical structure information during cross-modality image synthesis.

Main Methods:

• Learns a common feature encoding across modalities to create a shared latent space.
• Ensures consistent anatomical information between original and synthesized images.
• Preserves transformations between images within a domain through their cross-domain syntheses.

Main Results:

• Successfully applied to cross-modality skull and cardiac substructure segmentation.
• Demonstrated superior performance compared to state-of-the-art cross-modality segmentation methods.
• Effectively preserves high-level semantic and anatomical information.

Conclusions:

• The proposed anatomy-regularized approach enhances cross-modality image synthesis for medical segmentation.
• This method effectively mitigates the limitations of existing techniques in preserving semantic information.
• Offers a promising solution for improving medical image segmentation with limited labeled data.