Bridging MRI Cross-Modality Synthesis and Multi-Contrast Super-Resolution by Fine-Grained Difference Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-modal magnetic resonance imaging (MRI) faces challenges in accurately modeling inter-modal differences for imputation and reconstruction tasks.
• Existing methods sparingly address fine-grained differences arising from spatial misalignment and modality-specific structural variations.

Purpose of the Study:

• To integrate cross-modality synthesis (CMS) and multi-contrast super-resolution (MCSR) into a unified framework for improved multi-modal MRI analysis.
• To accurately model and address spatial misalignment and structural distinctions between MRI modalities.

Main Methods:

• A composite network architecture featuring a label correction module, CMS module, SR branch, and a difference projection discriminator.
• Adversarial learning with distinction-aware incremental modulation for controlled generation in the SR branch.
• Integration of deformable convolutions to handle cross-modal spatial misalignment at the feature level.

Main Results:

• The proposed approach effectively balances structural accuracy and realism in multi-modal MRI.
• Demonstrated overall superiority over state-of-the-art methods in comprehensive evaluations for both CMS and MCSR tasks.
• Experiments on three public datasets validated the framework's effectiveness.

Conclusions:

• The unified framework successfully addresses the challenge of fine-grained inter-modal differences in multi-modal MRI.
• The method offers a significant advancement for MRI reconstruction and imputation tasks.
• The developed approach provides a robust solution for enhancing the quality and accuracy of multi-modal MRI data.