Whole Brain 3D MR Fingerprinting in Multiple Sclerosis: A Pilot Investigation

Background: MR ngerprinting (MRF) is a novel imaging method proposed for the diagnosis of Multiple Sclerosis (MS). This study aims to determine if MR Fingerprinting (MRF) relaxometry can differentiate frontal normal appearing white matter (F-NAWM) and splenium in patients diagnosed with MS as compared to controls and to characterize the relaxometry of demyelinating plaques relative to the time of diagnosis. Methods: Three-dimensional (3D) MRF data were acquired on a 3.0T MRI system resulting in isotropic voxels (1x1x1mm 3 ) and a total acquisition time of 4min 38s. Data were collected on 18 subjects paired with 18 controls. Regions of interested were drawn over MRF-derived T 1 relaxometry maps encompassing selected MS lesions, F-NAWM and splenium. T 1 and T 2 relaxometry features from those segmented areas were used to classify MS lesions from F-NAWM and splenium with T-distributed stochastic neighbor embedding algorithms (T-SNE). Partial least squares discriminant analysis (PLS-DA) was performed to discriminate NAWM and Splenium in MS compared with controls. Results: Mean out-of-fold machine learning prediction accuracy for discriminant results between MS patients and controls for F-NAWM was 65% and approached 90% for the splenium. There was signicant positive correlation between time since diagnosis and MS lesions mean T2 (p=0.015), minimum T1 (p=0.03) and negative correlation with splenium uniformity (p=0.04). Perfect discrimination (AUC=1) was achieved between selected features from MS lesions and F-NAWM. Conclusions: 3D-MRF has the ability to differentiate between MS and controls based on relaxometry properties from the F-NAWM and splenium. Whole brain coverage allows the assessment of quantitative properties within lesions that provide chronological assessment of the time from MS diagnosis.


Introduction
Multiple Sclerosis (MS) involves a wide spectrum of neurological symptoms resulting in challenging clinical management based on symptomatology alone (1). Magnetic Resonance Imaging (MRI) has emerged as a powerful tool in the assessment of MS (2) with the requirement that imaging is performed using standardized imaging protocols (3).
With high diagnostic sensitivity, conventional MRI is able to describe disease dissemination in time and space (4), classify MS subtypes (2) and evaluate treatment response (2,5). However, conventional MRI may be limited when distinguishing ongoing in ammatory demyelinating pathology in normal-appearing white matter despite known disease processes (6) as well as functional disability. (7) It has previously been demonstrated that MS pathology can be described through quantitative spatial mapping of MRI-derived relaxometry parameters, such as longitudinal (T 1 ) and transverse (T 2 ) relaxation times or proton density (PD) (8). Further, parametric mapping may overcome the aforementioned limitations associated with MS diagnosis and staging by improving diagnostic accuracy (6,9,10) and predicting patient functional impairment (11,12).
MRF is a novel MRI technique that allows quantitative mapping of T 1 , T 2 and PD using acquisition schemes followed by matching of the data to synthetically generated signals. The details of MRF have been described previously (13) and involve the repeated acquisition of image data over a time course in which acquisition parameters such as the ip angle, pulse repetition rate (TR) and echo time (TE) are intentionally modi ed (14). Because the resultant time evolution of the signal in a given voxel is unique for a certain combination of tissue MR properties such as PD, T 1 and T 2 , MRF derived estimates of these parameters are generated by comparing the signal evolution history of a given voxel to a dictionary of pre-simulated signal evolutions (15). Acquiring brain relaxometry values in clinically feasible times in patients with MS have been proposed with the QRAP-MASTER pulse sequence (16). This technique has yet to meet the requirement of being obtainable within a relatively short acquisition time and as a result has had limited application as part of a standard, time constrained clinical MR examination. MRF has the potential to address this time constrain and has been described as a promising classi er of MS subtypes (17) within imaging times of several minutes. However, in that work the acquisition involved only 2D data and did not provide full brain coverage, limiting the clinical use of such an approach.
The purpose of this study is to determine if MR relaxometry maps derived from a fast 3D-MRF executed as part of a standard clinical MR examination sequence can differentiate frontal lobe normal appearing white matter (F-NAWM) and splenium in patients with MS versus healthy volunteers based solely on MRFbased relaxometry differences. Further, we hypothesize that MRF can detect MS lesions and establish a temporal relationship between relaxometry values and the time since diagnosis.

Image Acquisition and Reconstruction
All clinical data were acquired on two 3T MR scanners (Discovery MR750 and Discovery MR750W, GE Healthcare, Waukesha, WI) using an eight channel receive-only RF head coil. MRF data acquisition was performed using a 3D steady state free precession (SSFP) sequence with a multi-axis spiral trajectory (18). Adiabatic inversion pulses were used before each acquisition. The ip angle ramped schedule ranged from 0.778 to 70 degrees. Sequence details can be found in (15) and (19). The acquisition FOV was 25.6 x 25.6 x 25.6 cm 3 with 1 mm isotropic voxel resolution. The total acquisition time for the whole brain volume was 4 min 38 s. The T 1 range for the dictionary was from 10 ms to 3000 ms and T 2 from 10 ms to 2000 ms. Fingerprint reconstruction and dictionary matching were performed o ine using Matlab (Mathworks, Natick, Massachusetts) on a 64bit Linux workstation equipped with two 8-core Intel Xeon Gold 6244 CPU @ 3.60 GHz, 376 GB system memory, and NVIDIA Tesla V100 GPU. The reconstruction pipeline has been described elsewhere 16 .

Patient population
An Institutional Review Board (IRB) approved protocol was used to obtain MRF data in patients scheduled for a clinical MR exam. Informed consent was obtained by all the participants. All methods were carried out in accordance with institutional guidelines and regulations. The MRF sequence was acquired during the clinical MRI prior to the administration of a gadolinium-based contrast agent. A total of 18 subjects with an established diagnosis of MS were included: 14 subjects had relapsing remitting MS, 3 had secondary progressive MS and 1 had primary progressive MS. Three subjects had active gadolinium enhancing MS lesions. In the control group, 18 subjects were selected and paired to age and gender for each individual with MS.

Regions of Interest Analysis
Segmentations were performed manually using 3D-Slicer software (20) as described in Fig. 1. Four to ten lesions were selected for each patient with MS, with a total of 105 lesions across 18 patients, 10 of which were active lesions. Perilesional edema was not included in the segmentation of active lesions. Additionally, for each patient, one ROI each in F-NAWM and splenium of the corpus callosum were drawn.
F-NAWM was de ned as areas without signal changes on the standard T2 weighted images in the clinical exam. In the control group, corresponding ROIs were drawn in the F-NAWM and splenium. First order statistics (interquartile range, skewness, uniformity, median, energy, robust mean absolute deviation, mean absolute deviation, total energy, maximum, root mean squared, 90 percentile, minimum, entropy, range, variance, 10 percentile, kurtosis, mean) obtained from each ROI were analyzed. All segmentations were reviewed by a Board certi ed neuroradiologist.

Statistical Analysis
Given the nature of the data, paired analysis test of statistical signi cance comparing regions between cases and controls was performed. Speci cally, univariate comparisons were made for all individual features using non-parametric Wilcoxon signed-rank tests. Partial least squares discriminant analysis (PLS-DA) with repeated cross-validation (n = 5) was performed to discriminate F-NAWM and splenium in MS compared with controls, combining features from T 1 and T 2 relaxometry. Also, rst order statistics features were used to classify MS lesions from F-NAWM and splenium with T-distributed stochastic neighbor embedding algorithms (T-SNE). Classi cation was done combining T 1 and T 2 and for each separately.

Results
Twelve of the 18 MS patients (mean age of 49 ± 13 years (mean ± SD)) were female. In the control group (n = 18; age mean ± SD age: 49 ± 14), 12 patients were female. Mean time since MS diagnosis was Page 5/12 83 months. A representative MRF-based T 1 map paired with conventional imaging weighted imaging is shown in Fig. 2. Figure 3 describes the distribution of paired differences for each measurement combining T 1 and T 2 relaxometry map features. Partial least squares discriminant analysis (PLS-DA) for frontal NAWM and splenium is displayed in Fig. 4. Repeated cross validation (n = 5) showed mean out-of-fold accuracy = 65% for discriminant results between patients and controls for frontal NAWM, but mean out-of-fold accuracy approaching 90% for Splenium.
The T-SNE Plot for classi cation of MS lesions is displayed on Fig. 4. AUC analysis for selected features demonstrated that median and mean T 1 and T 2 allowed perfect discrimination (AUC = 1) between splenium and lesions for both T 1 and T 2 . Also, discrimination from F-NAWM was excellent (AUC = 1) and (AUC = 0.98) using median and mean for T 1 and T 2 , respectively. Table 1 lists T 1 and T 2 relaxometry ranges for all structures analyzed.  Table 2 lists the data used to identify if any relaxometry features were associated with time since diagnosis by all ROIs segmented. The 5 strongest rank correlations with time-since-diagnosis appear to largely correspond to T 2 mean lesion measurements as show on Table 2.

Discussion
This work describes the use of a novel whole brain 3D MRF sequence (18,21) in differentiating F-NAWM and splenium in patients with MS based on relaxometry estimates. Given the highly reproducible and accurate information provided by MR relaxometry (21,27) the results of this study, and the isotropic whole brain coverage afforded by this technique, MRF has the potential for use in the diagnosis of patients with MS. The previously described application of MRF in the normal brain (21,22), brain tumors (23,24), epilepsy (25) and Parkinson disease (26), suggests that MRF has the potential for even broader application beyond MS.
MRI currently is a fundamental clinical tool when guiding therapy for patients with MS (28). Given the complexity of the condition, several studies have been conducted with more advanced MRI techniques (such as myelin water fraction or functional MRI) to predict whether MS could be diagnosed by machine learning techniques (29)(30)(31)(32). Although the mentioned investigations have been successful, those techniques differ from MRF in that they do not provide a multi parametric approach from a single acquisition leading to lengthier exam acquisitions. Furthermore, the reproducibility of said techniques is not as well established as MRF-based relaxation estimates for both in vivo and phantom experiments (27,33).
F-NAWM demonstrated longer relaxation in patients with MS in our study. This has been described by other quantitative imaging investigations (34). Those changes are thought to be related with myelin histological changes in the white matter poorly de ned by imaging (35) and importantly could predict clinical disability (12). In this study, F-NAWM differentiation using MRF relaxation properties between cases and controls was fairly weak (mean out of fold accuracy = 65%). Given the moderate sample size, it is possible larger samples could describe more robust differentiation. Also, it is important to note prior studies (34,36,37) have provided estimates of the entire NAWM through the brain, potentially including areas adjacent to MS plaques that can have subtle signal changes. In order to avoid this pitfall, values stated in this work were from segmented areas that only included white matter with no changes in the conventional T 2 weighted imaging and double inversion recovery.
Splenium is partially responsible for interhemispheric connections within the brain (38). As such, studies describing splenium changes in patients with MS (39) have focused on diffusion tensor imaging.
However, histological changes in MS may also be responsible for relaxation lengthening in the splenium (34). The accuracy described in this study for classifying disease and control at this anatomical site based solely on MRF-based relaxometry changes was fairly strong (= 90%), identifying a major advantage of MRF. Given its potential to depict changes that are currently not seen or described in clinical practice, MRF may useful, especially in those cases where the diagnosis of MS is not clearly established by more conventional well established imaging protocols.
It is known that timing since diagnosis in MS can in uence normal tissue relaxation (7,11,37), and that those changes could predict clinical disability (40,41). Papadoulos et at (7) described NAWM relaxation changes in a longitudinal study covering 5 years. However, Davies et al (11) found no signi cant differences in a three year longitudinal study after accessing T 1 quantitative changes through NAWM and GM. In this study, T 2 lengthening was observed in MS plaques on those patients with the longest time from diagnosis of MS to imaging. These ndings could be related to a higher degree of Wallerian degeneration (42) although this nding has questionable clinical signi cance. Also, given this study was cross sectional, it would be valuable to investigate MRF through the same protocol in a longitudinal basis, so NAWM and splenium changes may be described and the faster acquisition as compared with the protocols mentioned (7,40,41) is a valuable tool for clinical application.
This study has several limitations. The relatively small sample size may not be su cient to effectively establish F-NAWM and splenium changes in MS as compared to controls. Also, F-NAWM segmentations represented a minimal fraction of the overall WM in all the patients included. Both active and non-active lesions were included, as de ned by gadolinium enhancement in conventional T1 weighted imaging, but given only 10 lesions were active, this study was not powered to detect changes within relaxometry for classifying lesion activity. Future studies with larger sample sizes and volumetric segmentation through normal appearing white matter may be considered. Author contributions: T.RM has made substantial contribution to the conception, design, acquisition, analysis, interpretation of data, drafted the work and substantively revised it. A.P. has made substantial contribution to the analysis, interpretation of data and substantively revised it. NG.C. has made substantial contribution to the analysis, interpretation of data and substantively revised it. RJ.W. has made substantial contribution to the interpretation of data and substantively revised it. Y.S. has made substantial contribution to the conception, acquisition and substantively revised the work. A.L. has made substantial contribution to the conception and substantively revised it. KP.M. obtained institutional grant for this research and has made substantial contribution to the conception, design, acquisition, analysis, interpretation of data and substantively revised it. Figure 3 Frontal NAWM and Splenium classi cation for MS compared to control: Partial least squares discriminant analysis between patients (cases) and controls for splenium and frontal NAWM (normal appearing white matter) within all 18 MS patients and 18 controls.