Quantitative MR Fingerprinting for Synthetic CT Via a Deformably-Registered Fusion Transformer Network
Abstract
Purpose
To develop and validate a deep learning framework for high-accuracy pseudo-CT synthesis from rapid, multi-parametric Magnetic Resonance Fingerprinting (MRF) for liver radiotherapy. This work addresses the technical challenge of translating quantitative tissue property maps into electron density information while ensuring spatial fidelity through integrated motion compensation.
Methods
We present a Deformably-Registered Fusion Transformer (DRFT) network for MRF-to-CT synthesis. The DRFT architecture uniquely combines two key components: 1) a cascaded transformer-based fusion encoder that jointly processes T1, T2, and proton density (PD) maps to extract complementary features, and 2) a trainable, diffeomorphic registration sub-network that compensates for respiratory motion between MRF and planning CT scans within the training pipeline. The model was developed using a cohort of 24 hepatocellular carcinoma patients. Performance was benchmarked against U-Net and pix2pix cGAN baselines using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error in Hounsfield Units (MAE).
Results
The DRFT framework achieved state-of-the-art synthetic CT quality, yielding an SSIM of 0.914 ± 0.011, a PSNR of 28.55 ± 1.31, and an MAE of 63.2 ± 7.8 HU, outperforming all baseline models. Multi-parametric MRF input was essential; the full T1+T2+PD model significantly surpassed the best single-parameter (PD-only SSIM: 0.885) and dual-parameter models (best dual-input SSIM: 0.894). Integrating the trainable registration module was critical, improving SSIM by 0.028 compared to an unregistered variant. Visual analysis confirmed enhanced preservation of bone anatomy and soft-tissue interfaces.
Conclusion
This study demonstrates the efficacy of a transformer-based fusion network with embedded deformable registration for generating synthetic CT from quantitative MRF data. The results confirm the necessity of synergistic multi-parametric guidance and explicit motion modeling for anatomical accuracy in the abdomen. This framework provides a robust technical foundation for quantitative MR-only treatment planning for liver cancers.