Dual-domain latent alignment for MRI-to-CT synthesis via target-aware manifold learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed Tomography (CT) is crucial for diagnosis and radiotherapy but uses ionizing radiation.
• Magnetic Resonance Imaging (MRI) offers a radiation-free alternative, driving research into MRI-to-CT image synthesis.
• Existing methods often require fully paired data, limiting training diversity and scalability.

Purpose of the Study:

• To develop a robust MRI-to-CT synthesis method that overcomes limitations of paired supervision.
• To enhance anatomical structure preservation, particularly for bone-related details, in synthesized CT images.
• To leverage unpaired CT data effectively for improved synthesis quality and geometric fidelity.

Main Methods:

• Proposed DELA-Net (Dual-Encoder Latent Alignment Network) for target-aware latent manifold alignment.
• Introduced a geometric reference selection strategy to identify informative unpaired CT samples as latent anchors.
• Decomposed optimization into coupled alignment tasks (MRI-to-unpaired CT, paired-to-unpaired CT) using a latent space modulator for structural fidelity.

Main Results:

• DELA-Net demonstrated a favorable balance between synthesis quality and cross-modality task translation.
• Achieved significant preservation of bone-related structures in downstream task evaluations.
• Experiments on three benchmark datasets validated the method's effectiveness.

Conclusions:

• DELA-Net offers a promising solution for synthesizing CT images from MRI, addressing the need for diverse training data and anatomical accuracy.
• The method effectively utilizes unpaired CT data for improved geometric consistency.
• Publicly available code facilitates further research and application in medical imaging.