Reg-APGAN: Registration-Guided Anatomy-Preserving GAN for CT-to-MR Translation

Area of Science:

• Medical imaging
• Artificial intelligence
• Image processing

Background:

• Computed tomography (CT) offers wide accessibility but limited soft-tissue contrast compared to magnetic resonance (MR) imaging.
• Current CT-to-MR translation methods often sacrifice anatomical reliability for perceptual realism, causing geometric distortion and limiting utility.
• There is a need for methods that synthesize MR-like contrast from CT while preserving anatomical integrity for improved downstream applications.

Purpose of the Study:

• To introduce Reg-APGAN, a registration-guided, anatomy-preserving framework for synthesizing MR-like contrast from abdominal CT scans.
• To address the limitations of existing CT-to-MR translation methods by ensuring anatomical reliability and geometric consistency.
• To evaluate the performance of Reg-APGAN in synthesizing pseudo-MR images and assess their utility in downstream tasks like segmentation.

Main Methods:

• Developed Reg-APGAN, a framework that unifies abdominal CT and MR data into a coronal space and performs hierarchical rigid registration using skeletal and organ labels.
• Implemented 2D slice-wise translation on rigidly aligned 3D CT-MR volumes with structure-aware supervision to maintain anatomical topology.
• Evaluated the method on the full abdominal cavity, a challenging deformable region, using 114 patient CT and T1-weighted MR datasets.

Main Results:

• Reg-APGAN achieved a 0.51 dB PSNR improvement, the highest MS-SSIM, and reduced Mean Absolute Error (MAE) by 6-7% compared to CycleGAN under weak-alignment conditions.
• The framework demonstrated a 13-15% reduction in region-of-interest (ROI)-based intensity and distributional errors.
• Pseudo-MR images generated by Reg-APGAN enabled more anatomically coherent downstream segmentation compared to CT alone and baseline translation methods.

Conclusions:

• Reg-APGAN successfully synthesizes structurally consistent MR-like visualization from CT by coupling anatomical registration with contrast translation.
• The framework enhances anatomical reliability in cross-modality synthesis, supporting multimodal AI development.
• While promising, Reg-APGAN requires external validation and radiologist reader studies for clinical adoption and is not intended to replace diagnostic MR imaging.