Introduction

One major thrust in radiology today is image standardization with rapidly-acquired, quantitative, multi-contrast information. This is critical for multi-center trials, for the collection of big data and the use of artificial intelligence in evaluating the data. In the past few years new quantitative, multi-parametric MR imaging methods have been presented addressing this need^1-3. Here we present further improvements of a novel 3D Silent Parameter Mapping⁴ method in terms of image reconstruction and synthetic CT image conversion⁹. We also assessed the accuracy of the obtained parameter maps using the quantitative NIST/ISMRM phantom and illustrate healthy volunteer scans acquired at 1.5T and 3T.

Methods

The 3D Silent Parameter Mapping method is based on T1 and T2 magnetization preparation combined with a segmented Zero TE (ZTE) readout to acquire interleaved T1 and T2 weighted images similar to the QALAS method (cf. Fig. 1). The PD map is measured upfront using ZTE with a low ~1deg flip angle (FA) resulting in negligible T1 saturation⁴, which simplifies the parameter estimation to two parameter fitting (i.e. T1, T2) and reduces crosstalk artifacts. The 3D radial ZTE images are acquired and reconstructed with 2x radial oversampling, which doesn’t require extra scan time but permits depiction of the anatomy beyond the nominal field-of-view (FOV). The images were reconstructed using a Deep Learning framework similar to as described by Lebel⁵. Dictionary based parameter fitting was implemented using orthogonal matching-pursuit, with the {T1, T2} dictionary entries calculated based on spoiled-gradient-echo (SPGR) signal equations⁴. The 2xFOV reconstructed PD images were tested for synthetic CT image conversion using a 2D multi-task network⁶ trained on N=140 pairs of registered ZTE and CT head&neck images obtained from independent studies. The 3D Silent Parameter Mapping was implemented and tested on a 1.5T MR450w and a 3T MR750w scanner using a GEM HNU head array coil (GE Healthcare, Chicago, IL). Its accuracy was assessed using a quantitative NIST/ISMRM phantom (CaliberMRI, Boulder, CO, USA) and compared to two alternative 3D parameter mapping methods including MAGiC^2,7 and QTI⁸. All volunteer experiments were performed with written informed consent and approval by local ethics commission.

Results

Figure 2 summarizes the quantitative T1 (left) and T2 (right) assessment of 3D Silent Parameter Mapping (without DL image reconstruction, green), 3D MAGiC (blue) and QTI (orange) using the NIST/ISMRM phantom at 3T. The 3D Silent Parameter Mapping demonstrates accurate and precise quantitation in the relevant range of T1 and T2 values found in-vivo. Its T2 values show increasing standard deviation for values very far distant from the echo time of the T2 preparation pulse used (i.e. 80ms in this case) . Figures 3 and 4 illustrate quantitative PD, T1 and T2 parameter maps obtained without (top) and with (bottom) Deep Learning image reconstruction in healthy volunteers at 1.5T (res=1.2mm, 6min05s) and 3T (res=1.0mm, 8min44s). Apparently, Deep Learning significantly reduces image noise and to a lesser extend also improves image sharpness resulting in excellent T1 and T2 quantitative parameter maps clean of cross-talk artifacts. Figure 5 illustrates 2xFOV extended reconstructions of the PD image (top) obtained from the 1.5T volunteer shown in Fig. 3 together with the Deep Learning derived synthetic CT (bottom).

Discussion

The presented improvements boost the performance of 3D Silent Parameter Mapping in terms of SNR, image sharpness and field-of-view coverage, allowing robust, high resolution 3D parameter mapping of the whole head in acceptable scan times. The obtained PD images are perfectly suited as input for Deep Learning based synthetic CT image conversion⁹. Together this forms a powerful package for silent, quantitative neuroimaging and also MR-only radiation therapy planning in terms of synthetic CT conversion and organ-at-risk (OAR) segmentation. The presented 3D Silent Parameter Mapping is available as work-in-progress package for GE MR and PET/MR scanners.

Acknowledgements

This research is part of the Deep MR-only Radiation Therapy activity (project numbers: 19037, 20648) that has received funding from EIT Health. EIT Health is supported by the European Institute of Innovation and Technology (EIT), a body of the European Union and receives support from the European Union´s Horizon 2020 Research and innovation program.