Technical Development and In Silico Implementation of SyntheticMR in Head and Neck Adaptive Radiation Therapy: A Prospective R-IDEAL Stage 0/1 Technology Development Report

Objective: The purpose of this study was to investigate the technical feasibility of integrating the quantitative maps available from SyntheticMR into the head and neck adaptive radiation oncology workflow. While SyntheticMR has been investigated for diagnostic applications, no studies have investigated its feasibility and potential for MR-Simulation or MR-Linac workflow. Demonstrating the feasibility of using this technique will facilitate rapid quantitative biomarker extraction which can be leveraged to guide adaptive radiation therapy decision making. Approach: Two phantoms, two healthy volunteers, and one patient were scanned using SyntheticMR on the MR-Simulation and MR-Linac devices with scan times between four to six minutes. Images in phantoms and volunteers were conducted in a test/retest protocol. The correlation between measured and reference quantitative T1, T2, and PD values were determined across clinical ranges in the phantom. Distortion was also studied. Contours of head and neck organs-at-risk (OAR) were drawn and applied to extract T1, T2, and PD. These values were plotted against each other, clusters were computed, and their separability significance was determined to evaluate SyntheticMR for differentiating tumor and normal tissue. Main Results: The Lin's Concordance Correlation Coefficient between the measured and phantom reference values was above 0.98 for both the MR-Sim and MR-Linac. No significant levels of distortion were measured. The mean bias between the measured and phantom reference values across repeated scans was below 4% for T1, 7% for T2, and 4% for PD for both the MR-Sim and MR-Linac. For T1 vs. T2 and T1 vs. PD, the GTV contour exhibited perfect purity against neighboring OARs while being 0.7 for T2 vs. PD. All cluster significance levels between the GTV and the nearest OAR, the tongue, using the SigClust method was p < 0.001. Significance: The technical feasibility of SyntheticMR was confirmed. Application of this technique to the head and neck adaptive radiation therapy workflow can enrich the current quantitative biomarker landscape.


Introduction
Magnetic resonance imaging (MRI) plays a vital role in visualizing tissues which are often indistinguishably isointense on computed tomography (CT) 1 .This becomes especially important in the field of radiation oncology due to most malignancies being in the soft tissue where CT has lower contrast.Due to this superiority, systems designed for radiation therapy (i.e., the linear accelerator, or Linac) have integrated diagnostic-level MRIs to create the 1.5T MR-Linac 2 , allowing for simultaneous radiation therapy delivery and MRI acquisition for anatomically dynamic structures (i.e., the lung).Furthermore, due to the availability of imaging acquired at each treatment fraction, strategies to adapt the treatment delivery based on imaging changes have gained traction and is known as adaptive radiation therapy, or ART.However, unlike CT, the radiation therapy workflow in MRI is currently focused around qualitative rather than quantitative representations of tumor and healthy tissue using the onboard MRI system of the MR-Linac.Though easy to interpret visually, the arbitrary signal values associated with standard relaxation "weighted" images in MRI present hurdles to the standard radiation therapy workflow, not the least of which are that these images are not ideal to monitor treatment response since they do not provide consistent quantitative measurements which can be compared across the different fractions of treatment 3 .In head and neck cancer, quantitative MRI probing tissue T1 and T2 properties has recently been investigated for diagnosis 4 , assessment of treatment response 5 , and assessment of normal tissue damage 6 .However, on the 1.5T MR-Linac, T1 and T2 mapping has been limited to acquisition times of exceeding 3 minutes and 5 minutes, respectively 7,8 .Therefore, acquiring T1 and T2 maps alone will require upwards of 8 minutes and suffer from higher misregistration error due to the potential for patient adjustments intra-and inter-scan.This time requirement is critical on the MR-Linac where, typically, less than 10 minutes are available for elective imaging within the clinical workflow 9 , and this will only decrease as broader patient coverage is demanded with the recently introduced Comprehensive Motion Management (CMM) software 10 .
The company SyntheticMR (Linköping, Sweden) has developed a single scan time (i.e., <6 minutes) simultaneous multiparametric MRI acquisition sequence originally known as QRAPMASTER 11 and more regularly known as multi-dynamic multi-echo (MDME) on Siemens MRI scanners, MAGiC on GE scanners, and SyntAc on Philips scanner.Through their postprocessing software, SyMRI, these acquired images can be reconstructed to quantitative T1, T2, and proton density (PD) maps and derivative synthetic contrast maps such as T1/T2/PDweighted and inversion recovery (IR), e.g., fluid attenuated (FLAIR), phase-sensitive (PSIR), and short inversion time (STIR).The most common implementation of the sequence is the 2D-MDME [11][12][13] which acquires multiple 2D slices, often with slice thicknesses of 3 -6 mm while inplane resolution is often <2 mm.However, a 3D version of the sequence capable of 1 mm isotropic acquisitions (3D-QALAS) 14 , has been developed and has shown clinically acceptable quantitative accuracy and repeatability in a multi-center 15 and multi-vendor study 16 .More details of the MR physics, technical considerations, and pulse sequence design can be seen in the review article by Hwang et al. 2022 17 .
SyntheticMR has seen increasing usage for diagnostic imaging [18][19][20][21] , however limited investigation has been conducted in usage for radiation oncology [22][23][24][25] and even fewer studies have focused on the head and neck 26,27 .Further, no studies to the author's knowledge have evaluated the technical feasibility of the SyntheticMR sequence on MR-Simulation (MR-Sim) or MR-Linac devices.The potential application of SyntheticMR to the radiation oncology space has high promise for increasing the dimensionality available for treatment monitoring and optimal adaptive therapy decisions.Specifically, within the head and neck oncology space, due to the large number of organs-at-risk (OARs) near the target, advanced quantitative MRI techniques would be advantageous to characterize simultaneous normal tissue dose-responses and tumor control more effectively.Therefore, the purpose of this study is to investigate the technical feasibility of integrating SyntheticMR into the head and neck adaptive radiation therapy workflow on both an MR-Sim and MR-Linac scanner.This will be presented using the radiotherapypredicate studies, idea, development, exploration, assessment, and long-term study (R-IDEAL) framework, as recommended by the MR-Linac Consortium, completing Stage 0 (radiotherapy predicate studies) and Stage 1 (first time use) systematic evaluations 28 .

MRI Acquisition Parameters
To evaluate SyntheticMR across the radiation oncology department, MRI scans were performed on a 3T Siemens Vida MR-Sim scanner (Siemens Healthcare; Erlangen, Germany) and a 1.5T MR-Linac (Unity; Elekta AB; Stockholm, Sweden).For this study, the 2D-MDME sequence was acquired with acquisition parameters as shown in Table 1.These parameters were chosen to optimize the acquisition for different clinical scenarios on the MR-Linac including the possibility for lymph node evaluation 29 (coarse sequence) and required resolution for stereotactic radiation therapy precision 30 (fine sequence).Further, due to the non-isotropic acquisition of 2D-MDME, all scans were acquired in the axial / transverse orientation to best visualize structures in the head and neck at all points in the radiation therapy workflow 31,32 .

MR Phantoms Assessed
The American College of Radiology (ACR) large phantom was used to evaluate geometric distortion and accuracy.The CaliberMRI "ISMRM/NIST" Premium System Phantom Model 130 phantom (CaliberMRI; Boulder, CO) was used as a reference for NIST-traceable T1, T2, and PD values 33 .This phantom includes 14 vials for each metric (42 total vials) suitable for T1 values between 20 and 1724 ms, T2 values between 9 and 853 ms, and PD values between 5 and 100% at 20°C on a 1.5T MRI scanner.This was confirmed during routine quality assurance that the temperature inside the bore averaged 20°C with minimal fluctuations.Reference values are also provided by the CaliberMRI for 3T MRI scanners.To evaluate repeatability, two repetitions in phantom were performed in a test-retest method using a coronal slice orientation.Similarly, to evaluate reproducibility, this process was reproduced across both the MR-Sim and MR-Linac scanners.

Healthy Volunteer and Patient Description
Two healthy volunteers were imaged on both the MR-Sim and MR-Linac and assigned the designation of Volunteer 1 (20s-year-old male) and Volunteer 2 (20s-year-old female) for future comparisons.Additionally, a 50s-year-old male with American Joint Committee on Cancer (AJCC) 8 th Edition Stage I (cT2, cN0, cM0, p16+) human papilloma virus (HPV) positive of oropharynx, squamous cell carcinoma of the left tonsil was scanned on the MR-Sim to evaluate the potential of SyntheticMR in the radiation therapy workflow.

Data Collection and Image Processing
The images were processed using the SyntheticMR post-processing software, SyMRI (StandAlone 11.3.11,Linköping, Sweden) developed specifically for the 2D-MDME sequence.
The geometric accuracy was evaluated using the ACR phantom by measuring four equal radially spaced diameters and comparing them to the expected length of 190 mm to within ±2 mm 34 .For the phantom analysis, a circular region-of-interest (ROI) was created in each vial using 3D Slicer 35 (https://www.slicer.org/)using the synthetically reconstructed T2-weighted MRI and the values within each ROI were extracted for processing in the quantitative T1, T2, and PD maps.A margin of approximately 10% of the vial's diameter was left to account for ringing artifacts hindering accurate readings 36 .For the in vivo analysis, the parotid and submandibular glands were chosen for analysis as the primary OAR for salivary dysfunction and were contoured automatically using a deep learning algorithm in the Advanced Medical Imaging Research Engine (ADMIRE) research software (v3.48.4,Elekta AB, Stockholm, Sweden).Additional critical structures including the tongue, bilateral infraorbital lymph spaces, mandible, and bilateral masseter muscles were contoured by medical students in RayStation Research 12A R v13.1.100.0 (RaySearch Laboratories, Stockholm, Sweden).

Statistical Analysis
All relevant analysis concerning statistical methods are formulated using the guidelines for reporting Statistical Analyses and Methods in the Published Literature (SAMPL) 37 .All computational analysis was completed using Python 3.8.10.To evaluate quantitative parameter accuracy, for each vial inside the ISMRM/NIST phantom, the values inside the ROI were extracted and the mean and standard deviation were calculated and compared to manufacturer reference values.For all calculations, analysis was restricted to clinical ranges of T1 (250 -2000 ms), T2 (30 -300 ms), and PD (20 -160 pu) 38 across both 1.5T and 3T MRI devices.Lin's Concordance Correlation Coefficient (LCCC) was used instead of Pearson's r 2 to evaluate direct agreement to reference values instead of generalized linearity.For measuring bias between the measured values and reference values, the mean bias was determined, and Spearman's rank correlation coefficient (i.e., Spearman's ρ) was calculated to test for significant generalized correlations between bias and the magnitude of the reference value.A p-value of 0.05 was used for statistical significance.These trends were further visualized using a Bland-Altman plot 39 .Cluster analysis was performed using an elliptical envelope and a soft margin assuming 40% outliers.To assess the separation of the clusters for desired ROIs, the purity measure was used to describe the proportion of desired data points that are within the desired cluster compared to other ROIs.Statistical significance between desired ROIs was calculated using the SigClust method 40 with soft thresholding and 100 Monte Carlo iterations following mean centering and variance normalization.

Phantom Analysis
For the geometric accuracy / distortion analysis using the ACR large phantom, the MR-Sim and both sequences on the MR-Linac showed within-tolerance agreement with the 190 mm expected diameter for each of the four measurements.Further, lines drawn along the geometry grid lines were straight and showed no intra-phantom distortion.
When evaluating quantitative parameter accuracy, the Lin's Concordance Correlation Coefficient between the measured and phantom reference values was above 0.97 for both the MR-Sim and MR-Linac when analyzing for T1, T2, and PD as shown in Figure 1.The highest agreement was seen in the PD values (average LCCC = 0.995), followed by T1 (average LCCC = 0.987), then T2 (average LCCC = 0.985).There is no significant difference between the LCCC of the MR-Sim and MR-Linac.
Figure 1.Correlation plot between the measured and reference T1 (left), T2 (center), and PD (right) for the MR-Sim and both coarse and fine sequences for the MR-Linac using a test-retest protocol.Abbreviations: LCCC = Lin's Concordance Correlation Coefficient.
The mean bias between the measured and phantom reference values across both repeat scans was 4.22% for T1, 6.32% for T2, and 3.11% for PD for both the MR-Sim and MR-Linac as shown in Figure 2. The MR-Linac coarse sequence had the lowest average T1 bias at 0.19% while averaging 1.70% for T2 and 9.62% for PD.The MR-Linac fine sequence had the lowest average T2 and PD bias at 0.76& and 1.92%, respectively, while averaging 5.12% for T1.The MR-Sim averaged biases of 5.42% for T1, 7.35% for T2, and 5.62% for PD.When calculating the correlation between reference values and bias, only the PD values showed significant pvalues (all p < 0.001) with a Spearman rank correlation coefficient averaging -0.96 across the MR-Sim and both sequences on the MR-Linac.
Figure 2. Bland-Altman plot between the measured and reference T1 (left), T2 (center), and PD (right) for the MR-Sim and both coarse and fine sequences for the MR-Linac using a test-retest protocol.

In vivo Analysis
An example of the SyMRI post-processing generated quantitative and subsequent synthetically generated contrast maps on the head and neck cancer patient on the MR-Sim in Figure 3. Note, the weighted contrast maps can be adjusted to different TE and TR values to achieve seemingly unlimited contrast options.Further, the inversion recovery (IR) sequences can be adjusted in a comparable way through the inversion time (TI).For the head and neck cancer patient, the T1, T2, and PD values inside each ROI were extracted and plotted against each other in two dimensions as shown in Figure 4.The purity measure was used between points inside each contour to determine cluster separability between the tumor and neighboring OARs.For T1 vs. T2 and T1 vs. PD, the GTV contour exhibited perfect purity while being 0.7 for T2 vs. PD.All cluster significance levels between the GTV and the nearest OAR, the tongue, using the SigClust method was p < 0.001.

Discussion
As shown in this technical feasibility analysis, SyntheticMR has the potential to create a paradigm shift in how imaging via MRI is done in the adaptive radiation oncology workflow.
When comparing to a prior study collecting T1 and T2 measurements in parotid glands on 1.5T scanners 41 , the mean and standard deviation T1, 578.8 (67.9), and T2, 104.5 (11.7), agreed with the distribution shown in this paper.To the author's knowledge, this is the first study investigating SyntheticMR on either the MR-Sim or MR-Linac.SyntheticMR can generate inherently co-registered T1, T2, and PD quantitative maps along with synthetically generated weighted and inversion recovery images in under six minutes.This has the potential to replace current time-intensive sequences and complex registration techniques currently used in the radiation oncology workflow for more efficient biomarker-based adaptive radiation therapy, increasing the adoption of more specialized quantitative imaging biomarkers approaches 42 at high temporal density.Some previously studied examples in the head and neck which could be adopted due to the time savings of SyntheticMR include apparent diffusion coefficient (ADC) 43 , dynamic contrast enhanced (DCE) 44 , diffusion-weighted imaging (DWI) 45 , and more emerging biomarkers such as T1-rho 46 .Further, the geometric accuracy test was well within passing criteria providing sufficient confidence for radiation therapy setup, planning, and delivery.
Despite its promise, the current limitations of SyntheticMR include susceptibility to patient motion artifacts which would propagate to all generated maps, suppression of the blood signal causing black blood features, and ghosting artifacts due to flow sensitivity.Another simultaneous, multiparametric, technique which addresses some of these limitations and has been investigated by one prior group on the 1.5T MR-Linac 9 is magnetic resonance fingerprinting, or MRF.This approach utilizes a series of dynamic scan acquisitions with each dynamic inheriting a pseudo-random set of acquisition parameters, typically the TR, TE, and flip angle.This pseudo-random acquired series is then matched to a pre-computed dictionary of MRI signal generation and fit to the most similar trajectory given the type of pulse sequence used.In comparison to SyntheticMR, MRF can acquire data using either a spin-echo or gradient-echo based approach allowing it to encode diffusion 47 and other weightings important in adaptive radiation therapy for head and neck cancer.However, MRF requires large precomputed dictionary creation which may add prohibitive computational requirements.Further, MRF is only commercially available on Siemens MRI scanners providing only simultaneous T1 and T2 quantification at the time of this writing limiting its immediate clinical translation in comparison to SyntheticMR which is commercially available on Siemens, GE, and Philips.The future applications of SyntheticMR in the general radiation oncology workflow is wide.
Recent movements have suggested the transition to probabilistic target definition instead of the currently used uniform target definitions 49 which builds upon the ideas presented in dose painting 50 .SyntheticMR has the potential to become a valuable asset in this space due to its multiparametric quantitative input at the voxel level which will help to inform probabilistic target definitions and optimal dose painting strategies.Techniques such as Bayesian and spatial statistics may also be employed upon SyntheticMR output to assess treatment response, identify sub-volume boost regions, and create more robust tumor control probability (TCP) and normal tissue complication probability (NTCP) models.
The multiparametric quantitative input at the voxel level may also be used as additional input channels for advanced image segmentation algorithms.This approach may be done manually using spatial statistics approaches, or automatically using deep learning techniques to identify optimal boundaries for each desired ROI in the three-dimensional quantitative space (T1 vs. T2 vs. PD).These boundaries may be created using large cohort studies of healthy volunteers and those with malignancies to generate consensus clusters for each desired ROI.Additional features from the synthetically generated contrast maps, such as radiomics, may also be included for improved performance 24 .Similar research utilizing SyntheticMR to achieve these tasks in the brain for white matter, gray matter, and others has been successfully demonstrated and employed clinically as a product of SyntheticMR.Additionally, after these regions have been successfully determined and validated, they may be used to assess the error of clinical contours and quantify contour uncertainty in the presence of more homogeneous or heterogeneous ROIs.

Conclusion
SyntheticMR is a promising novel approach to consolidate the currently inhibitory time requirements for multiparametric MRI acquisition incorporating both anatomical scans and quantitative information.In this paper, we demonstrated the potential of SyntheticMR to enhance the radiation oncology workflow in the following ways: 1) achievement of multi-contrast anatomical information and quantification in a single scan acquisition time allowing for higher throughput or addition of more research sequences 2) superior simultaneous quantitative accuracy at higher spatial resolution compared to existing techniques, and 3) clinically acceptable repeatability, reproducibility, and spatial accuracy.These factors aligned with the high temporal resolution of the MR-Linac (i.e., 33 fractions for a head and neck cancer patient) will exponentially increase both the clinical and research efficiency of the MR-Linac inside the adaptive radiation therapy workflow as it is currently used allowing for more effective balancing of tumor control with reduction of normal tissue toxicity.

Figure 3 .
Figure 3. Demonstration of the SyMRI post-processing package offered by SyntheticMR for the head and neck cancer patient on the MR-Sim scanner.Shown here is an axial / transverse slice at the level of the parotid glands, medial pterygoid muscle, and masseter muscle.The same set of images are shown in the Supplementary Materials for Volunteer 1 on the MR-Sim (Figure 3-S1), Volunteer 2 on the MR-Sim (Figure 3-S2), Volunteer 1 on the MR-Linac coarse sequence (Figure 3-S3), and Volunteer 1 on the MR-Linac fine sequence (Figure 3-S4).

Figure 4 .
Figure 4. Above, the quantitative spatial maps of neighboring OARs for the patient on the MR-Sim overlayed on top of their respective synthetically generated contrast map.The gross tumor volume (GTV) is shown in red dashed circle.Note, this patient was wearing a bite block and a permanent retainer which caused significant signal dropout in the oral tongue as shown by the white regions.Below, a demonstration of the separation generated by SyntheticMR for differentiating tumor and neighboring healthy tissues.Additional quantitative cluster analysis for more additional OARs in healthy volunteers is shown in Figure 4-S1 of the Supplementary Materials.

Figure 5 :
Figure 5: One potential application of SyntheticMR in the radiation oncology workflow to quantitatively adapt treatment based on detectable normal tissue damage in the salivary glands: 1) MRI acquisition overview, 2) generation of quantitative maps and subsequent synthetic contrast maps from SyMRI, 3) longitudinal acquisition schedule with high daily temporal resolution, and 4) subsequent evaluation of the deviations in the SyntheticMR maps.Note, the pulse sequence diagram (adapted from Fujita et al. 2024 15 ) on the top left is for the 3D-QALAS sequence, not the 2D-MDME used in this study.Further, the figure on the bottom right is adapted from Heukelom and Fuller 2022 48 .

Table 1 .
Acquisition parameters across all scanners utilized in this study.