Purpose: To optimize a deep neural network (NN) that incorporates k-space subsampling pattern selection and pharmacokinetic (PK) parameter estimation for reconstruction of dynamic contrast-enhanced (DCE) MRI.
Methods: 2D Cartesian phase-encoding subsampling patterns in a 3D Cartesian acquisition are modeled as independent and identically distributed samples from a multivariate Bernoulli distribution which is jointly learned within a convolutional recurrent reconstruction NN. A PK parameter estimation layer which produces PK parameters from the extended Tofts (eTofts) model using linear least-squares fitting follows the reconstruction network. The PK parameters of 25 real DCE-MRI scans were used to generate simulated dynamic image time series using the eTofts model which were then used for network training and testing. Gaussian noise with CNR=20 was added to the simulated data. 15 scans were used for training, 5 for validation, and 5 for testing. Loss functions investigated included combinations of 1) image reconstruction L₂ loss and 2) PK parameter estimation L₂ loss. An undersampling acceleration rate of 9 was used in all experiments. PK parameter estimation accuracy was evaluated by root mean squared error (RMSE). Performance was compared to a dictionary learning-based (DL) method.
Results: The combined image reconstruction and PK parameter loss yielded the lowest RMSE for all 3 eTofts model parameters, followed by the image reconstruction loss alone, and finally the DL method which demonstrated the largest deviation from ground truth. The combined loss reduced the RMSEs by 38.5%, 54.5%, and 41.0% for Kᵗʳᵃⁿˢ, k(ep), and v(p), respectively, compared with the DL method, with corresponding improved SSIM of reconstructed images (.98 vs .69).
Conclusion: A NN that jointly optimizes image subsampling patterns and reconstruction yields images and quantitative model parameters from undersampled MR scans that outperform DL techniques. Different subsampling pattern parameterization strategies and application on real DCE data will be investigated in the future.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01 EB016079.
Not Applicable / None Entered.