Synopsis

Keywords: CEST & MT, CEST & MT

An optimal saturation transfer MR fingerprinting (ST-MRF) acquisition schedule is critical for efficient and accurate tissue parameter mapping. To optimize RF saturation-encoded MRF acquisitions to a minimal number of saturation scan parameters for magnetization transfer contrast (MTC) and chemical exchange saturation transfer (CEST) parameter determination, we developed an optimization framework using a support vector regression-recursive feature elimination (SVR-RFE) method. Bloch simulations and in vivo studies showed that the proposed optimization method outperformed the quantification accuracy compared to existing methods. The SVR-RFE-based optimization method allowed us to reduce scan time by ~50% without sacrificing reconstruction accuracy.

Introduction

Conventional magnetization transfer contrast (MTC) and chemical exchange saturation transfer (CEST) MRI approaches acquire a series of images at multiple frequency offsets^1-3. Sometimes, a highly sampled Z-spectra acquisition is required to correct field inhomogeneity and isolate multiple CEST effects, which is time-consuming. A recent study showed that a linear projection method with L1-regularization can reduce frequency offset acquisition schedules, thus reducing total acquisition time⁴. Another optimization approach using a deep-learning neural network was introduced for quantitative MTC MR fingerprinting (MRF), which improved reconstruction accuracy and acquisition efficiency^5,6. However, a multi-pool exchange model increases the number of tissue parameters to be determined and computational complexity. Here, we developed a support vector regression-recursive feature elimination (SVR-RFE) method to optimize saturation transfer MRF (ST-MRF) with a three-pool exchange model (free water, semisolid macromolecule, and amide protons), which was compared with least absolute shrinkage and selection operator (LASSO)^7,8. LASSO typically uses least-squared regression with L1 regularization and minimizes the distance from each observation, whereas SVR maximizes the possible margin between the targets. We accelerated imaging acquisition by reducing redundant information from dynamic ST-MRF images.

Methods

Training dataset was synthesized by solving three-pool Bloch equations with a pre-defined ST-MRF schedule with sufficient dynamic scans (93 dynamic scans, a reference schedule) and tissue parameter combinations: free water T₁ (T₁^w) and T₂ (T₂^w) relaxation times, semisolid macromolecular proton exchange rate (k_mw), concentration (M₀^m), and T₂ relaxation times (T₂^m), amide proton exchange rate (k_sw), concentration (M₀^s), and T₂ relaxation times (T₂^s), and B₀ shift parameters. The importance of each dynamic scan parameter was measured by recursively removal of samples from the reference ST-MRF signal. The ST-MRF schedule with various numbers of dynamic scans (#30, #40, #50, #60, and #70) were obtained from the ranked reference dynamic schedule. For tissue quantification, a recurrent neural network was trained with the reference schedule, pseudorandomized (PR) schedule, LASSO-optimized schedule, and SVR-RFE-optimized schedule (Fig. 1). In the digital phantom study, as illustrated in Fig. 2, the quantification accuracy was evaluated by calculating normalized root mean square error (nRMSE) values between ground-truths and estimated tissue parameters. Nine normal subjects were recruited for in vivo study. All subjects were examined with the approval of the institutional ethics committee, and written informed consent was obtained prior to the study. 3D ST-MRF images were acquired from a multi-shot TSE sequence with fourfold (2 x 2) compressed sensing acceleration in the K_y-K_z plane at 3T MRI. Dynamic ST-MRF images were acquired with varied frequency offsets (Ω), RF saturation strengths (B₁), RF saturation times (Ts), and relaxation delay times (Td).

Results and Discussion

First, the SVR-RFE-based optimization technique was evaluated on digital phantoms (Figs. 2 and 3). As expected, the reference schedule with 93 dynamic scans had the lowest nRMSE, while the pseudorandomized schedule had the highest nRMSE. The optimization methods were compared against each tissue parameter and quantification accuracy was reported in Fig. 3. Overall, a reduction of the number of dynamic scans resulted in increased nRMSE. Nevertheless, both LASSO and SVM-RFE still showed a high degree of accuracy in the quantification of tissue parameters. The in vivo tissue parameter maps reconstructed from various ST-MRF schedules are shown in Fig. 4. Overall, the structural similarity index measure (SSIM) values of the SVR-RFE were higher than those of the LASSO optimization method, while the nRMSE values of the SVR-RFE were lower than those of the LASSO optimization method in in vivo experiments (n=9) (Fig. 5). We observed higher performance of SVR-RFE on in vivo data presumably because SVR calculating the support vectors lies on the edge of each class (tissue parameter) to make decision and hence, reduces the risk of overfitting on unseen data. With the SVR-RFE optimization algorithm, 50 dynamic scans provided the best trade-off between quantification accuracy (nRMSE <0.05 and SSIM > 0.95) and scan efficiency (reduced total scan time to ~50%).

Conclusions

We proposed an optimization framework to improve quantification accuracy and efficiency for ST-MRF. The SVR-RFE-based sequence optimization was tested with digital phantoms and demonstrated on healthy volunteer human brains. The optimized ST-MRF framework could provide quantitative MTC and CEST parameter mappings within a clinically feasible scan time.

Acknowledgements

This work was supported in part by grants from the National Institutes of Health.