Cross-Domain Transfer Learning of Clinical Score-Based Diffusion Models for Enhanced Low-Field Preclinical MRI Reconstruction
Abstract
Purpose
Portable MRI (pMRI) systems provide a compact, cryogen-free, and infrastructure-light platform for point-of-care and in-house clinical and preclinical imaging without patient or animal transport. However, the low magnetic field strength of pMRI inherently limits image quality. Score-based diffusion models have demonstrated success for clinical MRI reconstruction from degraded data, but their translation to preclinical pMRI remains largely unvalidated due to cross-species anatomical mismatch and limited preclinical training datasets. This work evaluates a clinically pre-trained diffusion model fine-tuned on high-field mouse MRI data to improve reconstruction in low-field preclinical MRI, thereby increasing the accessibility of high-quality MRI using widely deployable systems.
Methods
A variance-preserving stochastic differential equation (VP-SDE) diffusion model with DDPM U-Net (34.4M parameters) was initialized from clinical 3T human knee MRI weights (fastMRI) and fine-tuned on T2-weighted mouse brain MRI (11.1T/7T) from 167 subjects for 15 epochs. Training combined diffusion score matching with supervised reconstruction loss for efficient adaptation. Four Low-field degradation conditions were simulated to mimic pMRI scenarios via physics-informed k-space degradation: Gaussian blurring (resolution loss), T2*-like decay (k-space exponential damping), complex Gaussian noise (2-7 dB SNR), and B0-induced phase corruption. Inference used predictor-corrector sampling with k-space data consistency enforcement (weight=0.3). Performance was evaluated on 10 held-out subjects across four degradation regimes with 11.1T references.
Results
The method achieved strong reconstruction quality across all degradation settings. For noise-only degradation, SSIM reached 0.758±0.050, PSNR =32.72±1.53dB, and NMSE =2.42±0.77%. Under combined noise+blur, SSIM=0.676±0.049 with PSNR=27.73±2.60 dB. Adding T2* decay yielded SSIM=0.626±0.054 and PSNR=25.37±2.58dB. Under the most challenging regime (noise+blur+T2+B0 degrading), reconstruction maintained SSIM=0.657±0.078, PSNR=24.43±2.31dB, and NMSE=19.25±1.56%. Qualitative assessment showed substantial noise suppression and anatomical structure recovery without gross artifacts.
Conclusion
Transfer learning of a large-scale clinical diffusion model enabled effective enhancement of preclinical low-field MRI under severe degradation, with meaningful reconstruction achieved after only 15 fine-tuning epochs.