Author: Daniel O Connor, Mary Feng, Hui Lin, Hengjie Liu, Xin Miao, Michael Ohliger, Jess E. Scholey, Ke Sheng, DI Xu, Wensha Yang, Yang Yang 👨🔬
Affiliation: UCSF, University of California, Los Angeles, Department of Radiation Oncology, University of California San Francisco, Department of Radiation Oncology, University of California, San Francisco, Department of Radiation Oncology, University of California at San Francisco, University of San Francisco, Department of Radiology, University of California, San Francisco, University of California San Francisco, Siemens Medical Solutions USA Inc. 🌍
Purpose: The scanning time for a fully sampled MRI is lengthy. Compressed sensing (CS) has been developed to minimize image artifacts in accelerated scans, but the required iterative reconstruction is computationally complex and difficult to generalize. Image-domain-based deep learning (DL) methods (e.g., convolutional neural networks, CNNs) emerged as a faster alternative but face challenges in handling k-space data for their incapability of capturing global frequency relationships, a problem amplified with non-cartesian sampling commonly used in accelerated MR acquisition. In comparison, implicit neural representations (INRs) can directly model continuous signals in the frequency domain. We develop a patient-specific generative-adversarially trained INR (k-GINR) for MR reconstruction in undersampled non-Cartesian k-space.
Methods: k-GINR consists of two stages: 1) Training on prior acquisition and 2) Patient-specific reconstruction optimization. For stage-1, the INR network is trained with the generative adversarial network on a diverse patient cohort of the same anatomical region supervised by fully sampled acquisition. For stage-2, undersampled k-space data of individual patients is used to tailor the prior-embedded INR for patient-specific optimization. The UCSF StarVIBE T1-weighted liver dataset (118 scans; training: validation: testing = 7: 1.5: 1.5) were employed for model evaluation. Fully sampled raw radial k-space signals in one motion bin (around 450 spokes) with 26 sensitivity coils were used for training. The k-space signals were uniformly undersampled for validation and testing to pursue 3x, 10x, and 20x accelerations. k-GINR is compared with the Deep Cascade CNN and a CS method.
Results: k-GINR consistently achieved superior PSNR, SSIM, and RMSE to the baselines with a further widened performance gap in extreme accelerations (20x). The reconstruction time using k-GINR, CS and Deep Cascade CNN are approximately 3 min, 4 min and <1 s.
Conclusion: k-GINR, a novel INR network with adversarial training, is developed for direct non-Cartesian k-space reconstruction with promising outcomes demonstrated.