Adoption of Hybrid MRI-Linac Systems for the Treatment of Brain Tumors: A Systematic Review of the Current Literature Regarding Clinical and Technical Features

Abstract

Background

Possible advantages of magnetic resonance (MR)-guided radiation therapy (MRgRT) for the treatment of brain tumors include improved definition of treatment volumes and organs at risk (OARs) that could allow margin reductions, resulting in limited dose to the OARs and/or dose escalation to target volumes. Recently, hybrid systems integrating a linear accelerator and an magnetic resonance imaging (MRI) scan (MRI-linacs, MRL) have been introduced, that could potentially lead to a fully MRI-based treatment workflow.

Methods

We performed a systematic review of the published literature regarding the adoption of MRL for the treatment of primary or secondary brain tumors (last update November 3, 2022), retrieving a total of 2487 records; after a selection based on title and abstracts, the full text of 74 articles was analyzed, finally resulting in the 52 papers included in this review.

Results and discussion

Several solutions have been implemented to achieve a paradigm shift from CT-based radiotherapy to MRgRT, such as the management of geometric integrity and the definition of synthetic CT models that estimate electron density. Multiple sequences have been optimized to acquire images with adequate quality with on-board MR scanner in limited times. Various sophisticated algorithms have been developed to compensate the impact of magnetic field on dose distribution and calculate daily adaptive plans in a few minutes with satisfactory dosimetric parameters for the treatment of primary brain tumors and cerebral metastases. Dosimetric studies and preliminary clinical experiences demonstrated the feasibility of treating brain lesions with MRL.

Conclusions

The adoption of an MRI-only workflow is feasible and could offer several advantages for the treatment of brain tumors, including superior image quality for lesions and OARs and the possibility to adapt the treatment plan on the basis of daily MRI. The growing body of clinical data will clarify the potential benefit in terms of toxicity and response to treatment.

Introduction and Background

Radiation therapy has a crucial role in both primary and secondary brain tumors. Brain metastases (BMs) are the most common intracranial tumors in adults, with an incidence of about 15% in cancer patients, ¹ that is steadily increasing with the improvement of diagnostic imaging and the introduction of effective systemic therapy that prolong the lifespan.

The mainstay of BM local treatment is represented by surgery and/or radiotherapy, mainly consisting in whole brain radiotherapy (WBRT) for decades but with a progressive conversion towards single-fraction stereotactic radiosurgery (SRS) or hypofractionated stereotactic radiotherapy (SRT), allowed by modern radiotherapy techniques. ²

High grade glioma, including glioblastoma (GBM), is the most common malignant primary brain tumor of the adults. ³ Standard of care treatment includes maximal safe resection followed by adjuvant radiotherapy (with concurrent temozolomide in case of glioblastoma) and adjuvant chemotherapy. ⁴

Radiation treatment planning requires the definition of a GTV (gross tumor volume) which is expanded to CTV (clinical target volume, volume surrounding the GTV at risk of localization of subclinical disease) which is then expanded to PTV (planning target volume) accounting for set-up uncertainties, to ensure that the CTV receives the prescribed dose. ⁵ Healthy organs adjacent to the target volumes, defined as organs at risk (OARs) are as well contoured to limit their exposure to ionizing radiations. Both tumors and OARs are subject to variations in size, shape and position due to the effects of treatment, physiological movement of the organs and set-up changes.

Modern radiotherapy techniques, such as intensity modulated radiotherapy (IMRT) allowed to provide a more conformal treatment, resulting in better OARs sparing.^6–9

On the other hand, the steeper dose fall-off offered by modern radiotherapy techniques increases the risk of target missing, therefore the integration of adequate imaging is crucial in every step of the therapeutic workflow, from the definition of target volumes and OARs to the assessment of appropriate treatment delivery through image-guided radiotherapy (IGRT).

In the last decades, contours definition has been performed on computed tomography (CT) scans acquired with patients in treatment position with the aid of immobilization devices. Simulation images from CT are sub-optimal in defining the target and OARs in the brain, therefore registration with other diagnostic imaging techniques such as magnetic resonance imaging (MRI) is mandatory. Nonetheless, diagnostic imaging is acquired with a different set-up and often several days before simulation CT and therefore might not accurately represent the actual situation.

The use of IGRT techniques with imaging acquisition before the radiotherapy session has allowed a partial compensation of positioning errors, ¹⁰ but to date they are mainly based on CT imaging (Kilovoltage-CT, Megavoltage-CT or cone-beam CT) with limited resolution.

MRI allows a better definition of soft tissues than CT ¹¹ and has a consolidated and fundamental role in the management of cerebral tumors. Furthermore, MRI allows to generate different signal sequences of the same anatomical structures and to evaluate early biological modifications. Multiparametric MRI (mpMRI) includes functional studies such as diffusion-weighted imaging (DWI) with the measurement of apparent diffusion co-efficient (ADC) maps, dynamic contrast-enhanced (DCE) MRI and spectroscopy, that could aid to guide adaptive radiotherapy, anticipate and detect tumor response during and after treatment, discriminate between progression, pseudoprogression and radionecrosis and set the basis for radiomic analyses.^12–14

The possible advantages of MR-guided radiation therapy (MRgRT) include improved definition of treatment volumes and OARs, detailed evaluation of the morphovolumetric changes of the target and the OARs, adaptive treatment during each treatment session, avoidance of invasive procedures such as the placement of fiducial markers. Another advantage is represented by the possibility to assess predictive parameters of treatment response and toxicity by means of radiomics analysis of mpMRI and DWI images.^15,16 All these factors might allow reduced dose to the OARs and/or dose escalation to target volumes with the potential of limiting toxicity and improving cure rates.^17,18

Recently, hybrid systems integrating a linear accelerator and an MRI scan (MRI-linacs) have been introduced, that allow IGRT guided by MRI on which treatment plan can be adapted and could potentially lead to a fully MRI-based treatment workflow. The implementation of MRI-linacs required the development of dedicated solutions such as large bore magnets to accommodate specific immobilization devices, flat table tops, specific coil configurations and optimized MRI sequences. ¹⁹

Moreover, multiple issues should be considered for the paradigm shift from CT-based RT to MRgRT, such as the management of geometric integrity in order to achieve an adequate geometric accuracy and the definition of synthetic CT models that estimate electron density, which is not provided by MRI and is necessary for treatment planning in order to optimize the dose distribution.^20,21

Magnetic field affects photon beam dose distribution as a result of the deflection of secondary electrons through multiple phenomena, including the electron return effect (ERE) in which secondary electrons exiting in the air return to the tissue under the Lorentz force with an increase in skin dose and electron streaming effect (ESE) that is observed at the interface between tissues with elevated density difference and can result in increased dose deposition.^22,23

Despite the potential of MRgRT for the treatment of primary and secondary brain tumors, most of the currently published papers involved its application for lesions of other districts. In this review we present the existing literature regarding the implementation and use of MRgRT for brain cancers.

Methods

We performed a systematic review according to the recommendations of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. The keywords “MRI linac” or “MR linac” or “MRI guided radiotherapy” or “MR guided radiotherapy” and “brain” or “brain metastases” or “glioma” or “glioblastoma” were searched independently by two reviewers without time restriction (last update November 3, 2022) in the following databases: Scopus, PubMed, EMBASE and Cochrane Library. A total of 2478 records was identified, a preliminary selection was performed on the basis of the title and the abstract of the articles. Duplicates were excluded, as well as papers not published in English and abstracts and posters without full text. The full text of the resulting 74 articles was analyzed by both the reviewers responsible of paper selection and papers not regarding the adoption of MRI-linac for the treatment of brain tumors were removed, and the resulting 52 papers were included in this review. Data were collected independently by the two reviewers responsible of paper selection and then re-evaluated jointly to limit the risk of bias. Although the nature of the included papers was heterogeneous as well as the reported outcomes, data were organized by topics considered relevant for the clinical practice (as schematized by the sub-sections of the Results and Discussion section) and included in tables on the basis of shared outcomes when available.

As this is a Review article with no original data, approval by a Review Board was not required.

Results and Discussion

Commercially Available MRI-Linacs

Up to date, two MRI-linac systems are commercially available: the low-field 0.35 tesla (T) ViewRay MRIdian ® (ViewRay Inc., Oakwood, USA) and the high-field 1.5 T Elekta Unity® (Elekta Unity, Elekta AB, Stockholm, Sweden).

The ViewRay MRIdian system (ViewRay Inc., Oakwood, USA) in its first version integrated a 0.35-T split superconducting magnet with three Co-60 heads, replaced in the most recent version by a 6 megavolt (MV) flattening-filter-free (FFF) linear accelerator. ²⁴ The system is equipped with a double-focus, double stack multi-leaf collimator (MLC), composed by 138 tungsten leaves with no additional jaws, offset by the width of half a leaf resulting in a spatial resolution of 4.15 mm. Using MRIdian, daily MRI is performed and registered with the reference MRI and, if deemed necessary, contours are deformably registered and adapted on the image of the day with the possibility of recontouring and the plan is then optimized and verified. After approval, position is verified and during treatment real-time imaging at 4 frames per second (FPS) in a sagittal plane are acquired and patient-driven beam gating can be performed based on boundaries defined at the physician's discretion. ²⁵

Elekta Unity® (Elekta Unity, Elekta AB, Stockholm, Sweden) is a system that includes a 70-cm wide bore Philips (Philips Healthcare, Amsterdam, the Netherlands) 1.5 T scanner coupled with a single-energy 7-MV FFF linear accelerator with a source-axis distance of 143.5 cm. Treatment is delivered with step-and-shoot IMRT through a 160 leaves MLC in the in-plane direction with a maximum field size of 57.4 × 22 cm with field defining diaphragms in the cross-plane and a dose rate of 425 MU/min at isocenter.^26,27 Monaco treatment planning software offers two different workflows to generate treatment plans on the daily MRI for Unity®: Adapt-To-Position (ATP) or Adapt-To-Shape (ATS). In ATP the reference plan is re-optimized with an isocenter shift determined from rigid translation-only co-registration of the daily and reference images, maintaining the optimization parameters of the original plan and modifying the shape and weights of the original beams. ²⁶ The ATS workflow allows for plan adaptation based on the new patient anatomy and the plan is optimized on the daily MRI. The contours registration can be rigid or deformable and volumes and OARs can be modified and the adaptive plan can be comprehensively optimized from fluence map and adjusting as well the dosimetric objectives. ²⁶ During treatment cine-MRI can be acquired on the cardinal planes at 5 FPS, but gating has only recently been implemented.

MRI Sequence Optimization for MRI-Linac

The optimal MRI for RT planning should provide images with uniform signal intensity, lack of artifacts and high geometric accuracy.

Geometric accuracy of MRI is crucial for its application in RT, particularly in the context of an MRI-only workflow, as also small distortions may result in significant under-dose especially in low volume targets. ²⁷ Spatial distortion is influenced by multiple physical phenomena, including homogeneity of the field, gradient non-linearity (that increases with the distance from the isocenter), shimming control and patient-induced susceptibility effect (determined by the disruption of the field uniformity due to the patient presence, remarkably at the interfaces between the soft tissue and bone or air). ²⁸

The quantification and correction for geometrical distortion is complex, as it depends on multiple factors including the type of MRI machine, the adopted sequence and parameters, the strength of the field and gradient type and co-registration algorithms. An adequate management of this factors and the implementation of dedicated sequences, correction algorithms and quality assurance programs is thus mandatory, notably for targets with low volume or located peripherally or close to air cavities and for highly conformal treatments, and should reduce geometric distortion below 1 mm.^20,29

Specific algorithms allow to partially correct the impact of gradient-non-linearity and non-uniform intensity on geometric distortion. For instance, Bagherimofidi et al ³⁰ developed a fast distortion correction algorithm that allowed to reduce the distortion to 0.2 to 1.2 mm, based on tumor location. Ten patients were MRI-scanned for SRS of brain lesions using a 1.5-T scanner and an optimized pulse 3D T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) sequence; images were processed with a dedicated algorithm for patient- and machine-specific distortion; the validity of this model was confirmed with tests on phantoms. Inter-individual variability in human head shape and tissue composition resulted in heterogeneous changes in magnetic susceptibility, the most distorted areas and residual distortion after correction were localized around air cavities (eg sinuses).

The group of Taghizadeh ³¹ optimized MRI pulse sequences to limit distortion for an MRI-based workflow for SRS of metastatic brain lesions, using commercially available phantoms and images from 8 patients treated with Leksell Gamma Knife® Icon™. The uncorrected MR images showed a minimum–maximum deviation of 0.4 to 1.2 mm, comparable with that of CT scan, that was further reduced to 0.3 to 1.1 mm after geometrical correction.

To be integrated in the clinical workflow, MRI sequences acquired with MRI-linacs should be optimized taking into account the characteristics of the hybrid scanner and in order to reduce scanning time, while maintaining an image quality comparable with reference diagnostic MRI. In an evaluation of 59 patients enrolled in the MOMENTUM study with primary or secondary CNS tumors (including 20 high-grade glioma patients) treated on Unity, ³² DWI sequences were acquired daily during treatment on the MRI-linac and with the diagnostic 1.5 T scanner Philips Ingenia® at planning and at fractions 10 and 20 and one month after RT end. ADC measurements for target volumes and normal tissue obtained with the hybrid system were comparable with those obtained from the diagnostic scanner. The agreement on median ADC values between the two scanners over the GTV/CTV and brain regions was satisfactory, with small negative biases (maximum −9%) in white matter, grey matter, GTV, and CTV and substantial difference for CSF (maximum −20%); values of the bias could be used to adjust ADC values for direct comparison.

Chemical exchange saturation transfer (CEST) MRI offers metabolic information with a high sensitivity, but is usually performed with MRI field strengths of 3 T or more. Chan et al ³³ implemented CEST on the 1.5 T hybrid MRI-linac: repeated images were acquired from 54 patients with CNS tumors (mainly GBM – 51.9%, astrocytoma – 18.5% and resected BM – 13%) receiving radiotherapy on Unity®. Performing CEST imaging was feasible within the routine clinical workflow and the obtained images demonstrated significant contrast between tumors and normal white matter and the possibility to differentiate high and low grade lesions and monitor signal changes over time.

Strategically acquired gradient echo (STAGE) sequences, that allow the acquisition of several quantitative and qualitative datasets in short times, have been effectively implemented on a 0.35 T MR-linac, ³⁴ had good agreement with reference data and were sensitive to local tumor change (weekly during treatment and ∼ 2 months post-treatment) in 4 patients affected by primary brain cancers (3 GBM and one oligodendroglioma) treated with MRIdian system.

A novel approach to quantitative MRI that allows simultaneous efficient measurement of multiple tissue properties with a single acquisition is magnetic resonance fingerprinting (MRF). Bruijnen et al ³⁵ demonstrated the feasibility of MRF on a 1.5 T MRI-linac (MRL) achieving similar performances as those obtained on a diagnostic system, while Mickevicius et al ³⁶ investigate the feasibility of MRF on low-field (0.35 T) MR-linacs finding excellent reproducibility via phantom experiments.

While diagnostic MRI scans use advanced coils with optimized geometries to maximize image quality, daily MRI performed on MRL typically use flexible body coils combined with coil-specific frames in order to fit the patient's immobilization devices and this could limit image quality. Lu et al ³⁷ optimized whole-brain 3D-MRF with flexible body coils, demonstrating excellent repeatability and reproducibility with both values <3%, with acceptable geometric distortion and a 7% improvement in contour accuracy compared with MPRAGE T1-weighted imaging.

The conventional diffusion-weighted single-shot echo-planar-imaging (DW-ssEPI) technique suffers from limited resolution, severe distortion, and possibly inaccurate ADC at low field strength. Gao et al ³⁸ developed a reliable, accurate, and distortion-free diffusion sequence based on a turbo spin echo (DPTSE) acquisition that is practicable and specifically designed for assessment of tumor response to radiotherapy on an MRI-guided Tri-Cobalt 60 radiotherapy system. They found a mean distance difference between the locations of landmarks on the diffusion images and the reference CT images within 1.6 mm for DPTSE across 8 patients (2 with GBM and 6 with sarcoma), whereas this value was as high as 12 mm for DW-ssEPI; in the two GBM patients, ADC values estimated from DP-TSE sequence for CSF and white matter were within the ranges reported in literature.

Generation of Synthetic CT Scan (SynCT)

An MRI-only workflow could offer several advantages, including set-up and image registration error minimization and avoiding exposure to ionizing radiation from repeated CT scans. Since image voxel intensities of MRI are not associated with electron density of tissues as in CT, dose calculation on MRI requires a conversion from MRI to a synthetic-CT (synCT). The development of an adequate synCT model is complex, as MRI voxel intensity values are not standardized and vary with scanning parameters, manufacturer and magnetic field strength. Moreover, tissues with extremely different electron densities (EDs) might have overlapping MR intensity.

Different strategies have been implemented to generate synCT, including tissue segmentation based methods with bulk density assignment, estimation of the attenuation map using registration from a defined atlas, voxel based methods and machine learning techniques.^39–68 Generally, deep learning methods demonstrated superior than other techniques for the generation of synCTs and, once the model is trained, computational time to generate synCT is limited to a few seconds. The performance of these methods is dependent on image registration accuracy and limited by the heterogeneity of disease volume and geometry, particularly in case of large lesions and post-surgical modifications. Other possible sources of uncertainty are represented by different set-up and immobilization devices between repeated scans and variability in bone density among individual patients.

Multiple experiences demonstrated excellent geometric accuracy^39–44 and dosimetric agreement^{39,43,48–51,55–57,59,61,66} between simulation CTs and synCTs and the feasibility of adopting synCt to generate digitally reconstructed radiographs (DRRs) for IGRT.^{46,47,59,66,69} Larger errors were identified at air and bone interfaces^{44,55–57,67} possibly due to elevated intensity gradient and imperfect alignment between different imaging modalities.

Papers regarding generation of synCTs are summarized in Table 1 and a sample of a synCT generated for a glioblastoma patient is represented in Figure 1.

Table 1.

Report of Different Syntetic CT Generation Methods.

Author	Source	Generation method	Results	Limits
Uh et al 39	T2-weighted turbo spin-echo MRI	Atlas-based vs bulk density assignment	Atlas-based better accuracy than bulk density assignment; multiple atlases outperformed single-atlas models: 98% of the cases satisfactory dosimetric and geometric accuracy (<2% maximum dose and 2 mm range).
Boukellouz et al 40	T2-weighted MRI of 11 patients	Multi-atlas based	Average correlation between CT and synCT of 0.92 and DSC of 0.72 for bone.
Kim et al 41	T1-weighted and T2-weighted MRI of 20 brain cancer patients	Bulk-density assignment	DICE similarity coefficient (DSC) over 88% for all OARs with the exclusion of bone	Low similarity for bone (DSC 73.5%)
Sjölund et al 42	1.5 T T1-weighted MRIs of 10 patients that underwent stereotactic radiosurgery	Atlas-based regression	MAEs in the range of 113.4-125.6 HU
Demol et al 43	3 T 3D T1-weighted MRIs from 22 patients with 35 brain lesions treated with Cyberknife	Two methods (one based on deformation only and one that adopted actual MR intensities)	MAE < 50 HU for soft tissue. Adequate target coverage with dose differences between −2.9% and 3.1% and agreement within 2% for 85%	Higher discrepancies for low-density tissues and bones (up to 600 HU); high dosimetric differences for two patients with unusual anatomy
Yu et al 44	T1-weighted MRI sets of 25 brain cancer patients that underwent SRT or SRS	Tissue segmentation and voxel-based mapping	MAE for air, soft tissue and whole body 24 HU, 26 HU and -125 HU; contour distance of external skull < 1 mm	Major errors in complex structures such as mastoid sinuses.
Edmund et al 45	MRI datasets of 5 brain cancer patients	Six different voxel-based approaches	Satisfactory geometrical agreement, best results with regression. Mean dosimetric deviation between synCT vs CT within 2%, best agreement for statistical regression and threshold methods	highest dose discrepancies in the gradient dose region of the brainstem.
Weisinger et al 51	MRI datasets of eight patients with brain tumors	Tissue classification and regression	GPR 99.56% 2%/2 mm; dose difference to the PTV vs plans generated on CTs 0.23%
Price et al 46	MRI datasets from 12 brain cancer patients	Voxel-based weighted summation of 5 tissue classification	DRRs in close agreement with CT-generated (shift 0-0.7 mm).
Morris et al 47	MRI sets from 10 brain cancer patients	Voxel-based weighted summation	Average absolute shifts 0.77 mm and 0.76 mm for CT and synCT.
Li et al 55		Deep learning convolutional neural network (CNN)	EPID for MRL: 3D-γ passing rates 3%/2 mm of 97.42%
Liu et al 52	1.5 3D T1-weighted MRI (40 sets for training and 10 for evaluation)	Convolutional neural network	DSC of 0.94 for air, 0.94 for soft tissue and 0.85 for bone; MAE 75 HU. Dosimetric discrepancy 1.39% PTV Dmax and 0.27% for the PTV V95. Dosimetric constraints were respected for all OARs. Average global 3%/3 mm GPR 99.2%.
Massa et al 53	Multiple sequences (81 sets for training and 11 for evaluation)	Convolutional neural network	MAE for whole brain was in in a range between 44.56 and 51.24 HU
Lui et al 54	From MRIs of 20 brain cancer patients	Voxel-based tissue classification clinically-ready algorithm	Adoption of a modified scheme allowed a GPR 1%/1 mm of 92.2%; mean dose difference in PTV D95 and Dmax of −0.1% and −0.3%, no significant difference for any OAR (all within 0.2 Gy).
Paradis et al 55	MRI datasets of 12 glioma patients	Probabilistic voxel classification	Average MAE synCT vs CT 142 HU for the skull and 24 HU for the brain. Mean PTV difference CT vs synCT 0% for D95 and D5 and −0.2% for Dmax; average OARs difference 0%. GPR 96.4% at 2 mm/2%	discrepancies in close proximity of air cavities, although clinically insignificant.
Zheng et al 56	MRI sets of 10 brain cancer patients	Hybrid magnitude and phase MRI processing	Mean GPR 99.4% at 2%/2 mm; mean MAE synCT vs CT-SIM of 147.5 HU	larger errors at bone-air interfaces
Wang et al 57	MRI datasets of 19 brain lesions treated with SRS	Patch-based random forest method	Dose difference vs CT-based plans < 0.6% or 0.2 Gy for all PTV metrics and <0.02 Gy for all DVHs. GPR 3%/3 mm 99%	errors could be observed in small volumes around air cavities and bone.
Ranta et al 58	Dixon MRIs of 20 patients	Attenuation correction method	Dose comparison (Dmax, D0.1cc, D2%, D50%, D95%, D98%, Dmean) showed differences for PTV <0.4% and <1% for OARs; mean GPR 95.7% 2%/2 mm and 96.5% 1%/1 mm for glioma and metastasis, respectively.
Aouadi et al 59	1 and T2-weighted MRI of 13 brain cancer patients	Multi-scale dual-contrast patch-based method	MAE of 99.69 HU in bone vs CT; highest DVH metric deviation vs CT-based plans was 0.43% for PTV and 0.59% for OARs. DRRs shifts all <2 mm and <2° vs conventional DRRs.
Han 67	T1-weighted MRI from 18 brain cancer patients	Deep convolutional neural network algorithm	Superior than atlas-based approach: average MAE difference 84.8 HU vs 94.5 HU	larger errors at bone and air boundaries.
Lerner et al 60	Dixon MRIs form 20 patients with brain malignancies	Commercially available deep learning convolutional neural network	MAE within body contour of 62.2 HU. Plans generated on synCTs dose differences <0.2% for target, chiasma and brainstem; mean GPR 99.1% at 1%/1 mm.
Dinkla et al 61	T1-weighted gradient echo MRIs from 52 brain cancer patients	Dilated convolutional neural network algorithm	Mean dose deviation vs CT-base plans <1.5% inside PTV and <1% for mean and maximum dose for all the OARs; mean GPR 98.8% at 1 mm/1%; MAE within body contours 67 HU.
Gupta et al 48	T1-weighted MRI of 60 brain cancer patients	Deep learning U-Net architecture	MAE 17.6 HU in soft tissue; mean shift −0.1/−0.2 mm in all the axes vs CT	Mean PTV difference 2.3%, mainly for lack of immobilization mask in synCT
Kazemifar et al 49	T1 -weighted spin echo MRI of 77 brain cancer cases	Generative adversarial network	MAE 47.2 HU for bone; mean difference dose vs CT-based plans insignificant and <1% for all DVH parameters of PTV and OARs; MPR 99.2% and 94.6% (2%/2 mm and 1%/1 mm).
Wang et al 62	T1-weighted MRIs and CTs from 26 patients (training) and 5 used for validation	Generative adversarial network	Mean structural similarity indices between CTs and synCTs were 0.96 (cosine angle distance) and 0.84 (mean structural similarity).
Li et al 63	T2-weighted MRI datasets of brain cancer patients from two hospitals	Generative adversarial network	Best performance was achieved with a combined model using both source and target datasets, MAE of 74.56 HU.
Koike et al 64	580 pairs of multiple MRI sets for training, 15 GBM patients for evaluation	Generative adversarial network	Mean MAEs for the whole body, soft tissue and bone region were 108.1, 38.9 and 366.2 HU. D2%; small (<0.7%) but significantly higher dose to OARs in synCT plans. GPR 99.7% at 3%/3 mm, 98.9% at 2%/2 mm and 94.2% at 1%/1 mm.
Tang et al 65	27 pairs of T1-weighted MRI sets for training, 10 sets of brain cancer patients for evaluation	Generative adversarial network	Mean MAE 60.52; average GPR at 3%/3 mm and 2%/2 mm criteria 99.76% and 97.25%. Slight non-significant dose difference for target and OARs for plans on synCT (range −0.77/+1.33%).
Emami et al 50	single T1-weighted MRI	Generative adversarial network deep learning model	Outperformed a deep convolutional neural network algorithm, better representation of regions of abnormal anatomy.
Liu et al 66	T1-weighted postgadolinium MRI of 12 brain cancer patients	Generative adversarial network deep learning model	Mean difference ≤0.10 Gy PTV D95% and ≤0.13 Gy for OARs vs CT-based plans; mean GPR 99.9% and 99.0% at 2%/2 mm and 1%/1 mm; mean difference in CBCT-synCT registrations <0.2 mm

MAE, mean absolute error; HU, Hounsfield unit; GPR, gamma pass rate; DRRs, digitally reconstructed radiographs; EPID, electronic portal imaging device; DVH, dose volume histogram.

Figure 1.

Comparison of simulation CT (axial, coronal and sagittal scans—first line) and syntetic CT (second line) of a patient that underwent radical surgical resection of glioblastoma.

It has to be noted that in all the following papers synCTs were generated from diagnostic MRIs. Indeed, while some papers have been published for tumors of other sites, to the best of our knowledge no full text papers only a few abstracts and posters have been released regarding synCTs production from on-board systems for brain cancer.

In 14 patients with pediatric brain tumors ³⁹ atlas-based schemes using T2-weighted turbo spin-echo MRI performed better than bulk density assignment, and schemes using multiple atlases outperformed single-atlas models; in more than 98% of the cases achieved satisfactory dosimetric and geometric accuracy evaluation (<2% maximum dose and 2 mm range).

Boukellouz et al ⁴⁰ generated synCT from T2-weighted MRI of 11 patients using a multi-atlas based algorithm, achieving an average correlation between CT and synCT of 0.92 and a Dice similarity coefficient (DSC) of 0.72 for bone.

Kim et al ⁴¹ generated synCT of 20 brain cancer patients from T1-weighted and T2-weighted MRI with a bulk-density assignment method. Considering 8 structures (body, air, eyeball, lens, cavity, ventricle, brainstem, and bone) DSC was over 88% for all of them with the exclusion of bone (DSC 73.5%).

Sjölund et al ⁴² generated synCT from 1.5 T T1-weighted MRIs of 10 patients that underwent SRS adopting an atlas-based regression methods, obtaining a MAEs in the range of 113.4 to 125.6 HU.

Demol et al ⁴³ used 3 T 3D T1-weighted MRIs from 22 patients with 35 brain lesions treated with Cyberknife to generate synCTs with two methods (one based on deformation only—MRdef—and one that adopted actual MR intensities—MRint) and simulate treatment plans (at the same prescribed dose). Soft tissues presented a MAE < 50 HU, while discrepancies were higher for low-density tissues and bones (up to 600 HU), with significant differences for MRdef compared with CT and better results for MRint. Target volume coverage was adequate and differences of DVH parameters considered (D98, D95, D50, D05, D02, Dmean) between plans simulated on CTs and synCTs were between −2.9% and 3.1%, except for two cases (both with unusual anatomy) and agreed within 2% for 85% of patients.

Yu et al ⁴⁴ obtained synCTs from T1-weighted MRI sets of 25 brain cancer patients that underwent SRT or SRS on the basis of tissue segmentation and voxel-based mapping: mean absolute error (MAE) differences in density for air, soft tissue and whole body voxels were 24 Hounsfield (HU), 26 HU and −125 HU, respectively; average difference between the contour of external skull was 1 mm and mainly due to small differences in the positional angles, with major errors only in complex structures such as mastoid sinuses.

Weisinger et al ⁵¹ generated synCT through a conversion based on tissue classification and regression from MRI datasets of eight patients with brain tumors (three GBM, two meningioma, one craniopharyngioma, two metastases). Dice coefficient for bone overlap with CT scans was 0.73%, average dose difference to the PTV between plans generated on CTs and synCTs was 0.23%, with a gamma passing rate of 99.56% with 2%/2 mm.

Edmund et al ⁴⁵ compared six different voxel-based approaches (including bulk density assignment, Bayesian segmentation, threshold-based segmentation and statistical regression) to generate synCT from MRI co-registered with CT scan acquired with a thermoplastic mask of 5 brain cancer patients. A spherical 3 cm tumor target was simulated and treatment plans were generated on MRI or CT simulation to erogate 60 Gy in 30 fractions. Geometrical agreement was satisfactory for all the methods, with the best results achieved with statistical regression. The mean dosimetric deviation between plans calculated on CT and synCT was within 2%, with the best agreement for statistical regression and threshold based method and highest discrepancies in the gradient dose region of the brainstem.

In a study by Price et al, ⁴⁶ synCT generated adopting voxel-based weighted summation of 5 tissue classifications and using an anthropomorphic skull phantom and datasets from 12 brain cancer patients allowed to attain DRRs in close agreement with those generated from CT (registration shift difference from 0 to 0.7 mm).

Similarly, in the work by Morris et al ⁴⁷ DRRs reconstructed from synCT generated from MRI sets from 10 brain cancer patients (12 lesions, 7 postsurgical) using voxel-based weighted summation had satisfactory agreement with DRRs from CT scans, with a Dice similarity coefficient >0.95 and average absolute shifts were 0.77 ± 0.58 and 0.76 ± 0.59 mm for CT and synCT.

Lui et al ⁵⁴ adopted a voxel-based tissue classification clinically-ready algorithm to generate syCTs from MRIs of 20 brain cancer patients (12 with glioma, 6 with meningioma, 1 with chordoma and 1 metastasis). Treatment plans were recalculated with VMAT technique on synCT: while the initial dose differences and gamma analysis were sub-optimal, the adoption of a modified scheme with a specific calibration curve and the inclusion of a pseudo-skin (in order to compensate for the underestimation of body dimension detected in synCT) allowed to achieve a gamma passing rate 1%/1 mm of 92.2% and a mean dose difference in PTV D95 and D_max of −0.1% and −0.3% respectively and no significant difference for any OAR (all within 0.2 Gy).

Liu et al ⁵² developed a deep convolutional neural network approach, using 1.5 3D T1-weighted MRI from 40 non oncologic patients to train a model that was evaluated on 10 clinical cases of BM treated with SRT to generate synCT. Compared with co-registered kVCT images, the algorithm generated synCT with a DSC of 0.94 for air, 0.94 for soft tissue and 0.85 for bone and a MAE of 75 HU. Plans were generated on synCT at the same prescription dose (24-32 Gy in 3-8 fractions) with VMAT technique. The absolute percentage differences between plans simulated on CT and synCT were 0.24% ± 0.46% for PTV volume, 1.39% ± 1.31% for PTV D_max and 0.27% ± 0.79% for the PTV V95. Dosimetric constraints were respected in all the cases for the considered OARs (brainstem, chiasm, lens, optical nerves, cochleae) and average global 3%/3 mm Gamma analysis pass rates of 99.2% was recorded.

Massa et al ⁵³ adopted a convolutional neural network deep learning model on a training set of 81 patients to generate synCTs for 11 patients with brain cancer form four common clinical MRI sequences: the MAE for whole brain was in in a range between 44.56 HU for post-contrast T1-weighted gradient-echo and 51.24 HU for fast spin-echo T2-weighted FLAIR.

Li et al ⁶⁹ proposed a novel deep learning-based 3D in vivo dose reconstruction model using an electronic portal imaging device (EPID) for MRL: a pre-trained convolutional neural network (CNN) model yielded 3D-γ passing rates (3%, 2 mm) of 97.42% and MAE (%) of 0.88 for the brain.

Paradis et al ⁵⁵ produced synCTs from MRI datasets of 12 glioma patients with heterogeneous localization and volume (18-281 cm³) using probabilistic voxel classification, with an average MAE between synCT and CT of 142 HU for the skull and 24 HU for the brain. Mean differences between PTVs calculated from synCTs and simulation CTs were 0% for D95% and D5% and −0.2% for D_max; average difference of dose to the OARs was as well 0% and mean gamma passing rate was 96.4% at 2 mm/2%; larger dose discrepancies were observed in close proximity of air cavities, although clinically insignificant.

Zheng et al ⁵⁶ generated brain synCTs adopting a hybrid magnitude and phase MRI processing pipeline on MRI sets from 10 brain cancer patients, achieving a mean gamma passing rate of 99.4% at 2%/2 mm with acceptable plan quality and MAE between synCT and CT-SIM of 147.5 HU (with larger errors at bone–air interfaces).

Ranta et al ⁵⁸ re-calculated treatment plans on synCTs generated with attenuation correction method from Dixon MRIs of 20 patients (10 with glioma and 10 with metastases) who had previously received radiotherapy. Compared with plans generated from CT, parametric dose comparison (D_max, D0.1cc, D2%, D50%, D95%, D98%, Dmean) showed differences for PTV <0.4% and <1% for OARs; mean gamma pass rates were 95.7%; 2%/2 mm and 96.5%; 1%/1 mm for glioma and metastasis patients, respectively.

Patch-based random forest learning method demonstrated a good accuracy for the generation of synCT MRI sets of brain cancer patients ⁷⁰ and in some series was superior than atlas-based methods. ⁷¹

Lerner et al ⁶⁰ clinically evaluated a commercially available deep learning convolutional neural network algorithm on Dixon MRIs form 20 patients with brain malignancies (10 with metastases and 10 with glioma) of which 14 had areas of resected skull bone and re-calculated treatment plans (prescribed dose 24-60 Gy in 2-10 Gy per fraction). For all patients, including those with areas of bone resection, synCTs were successfully generated, with a MAE within body contour of 62.2 HU. Plans generated on synCTs resulted in dose differences below 0.2% for all the parameters considered (Dmean, D98, D2) for target volumes, chiasma and brainstem; mean gamma pass rate was 100% within PTV and 99.1% for the full dose at 1%/1 mm.

A patch-based random forest method was adopted also by Wang et al ⁵⁷ to develop synCT from MRI datasets of 14 patients (19 brain lesions treated with SRS at a dose of 18-21 Gy): compared with plans performed on simulation CT, dose difference in all PTV DVH metrics was <0.6% or 0.2 Gy, while all DVH parameters for selected OARs (brainstem, chiasm, optic nerves) were within 0.02 Gy; average gamma passing rate (3%/3 mm) was 99%. Residual errors were observed in small volumes around air cavities and bone.

Aouadi et al ⁵⁹ used a multi-scale dual-contrast patch-based method to generate synCT from T1- and T2-weighted MRI of 13 brain cancer patients, resulting in MAE of 99.69 HU and a Dice of 83% in bones compared with CT; agreement of treatment plans (60 Gy in 30 fractions) based on synCT with conventional plans was excellent as the highest DVH metric deviation was 0.43% for PTV and 0.59% for the OARs (eyes, optic nerves, brainstem). Accuracy of DRRs generated form synCT was satisfactory, as registration shifts were all <2 mm and <2° compared with conventional DRRs.

A deep convolutional neural network (DCNN) method developed by Han ⁶⁷ was applied on T1-weighted MRI from 18 brain cancer patients and resulted superior than an atlas-based approach: average MAE was 84.8 HU and was significantly better than that achieved by atlas-based method (94.5 HU); larger errors were reported at bone and air boundaries.

Dinkla et al ⁶¹ generated synCT through a dilated convolutional neural network (CNN) algorithm using T1-weighted gradient echo MRIs from 52 patients (26 patients used for training and 26 patients for evaluation) with heterogeneous brain tumors (38 metastases, 6 meningiomas, 3 pituitary gland adenomas, 3 vestibular schwannomas, 1 chordoma, and 1 hemangioblastoma; prescription dose 14-60 Gy in 1-30 fractions). Mean dose deviation between plans generated on synCT and CT was constantly <1.5% inside the PTV; deviations in mean and maximum dose were limited and well within 1% for all the OARs (brainstem, eyes, cochleae, chiasm, optic nerves, pituitary gland). Mean gamma passing rate was 98.8% at 1 mm/1% and MAE within the intersection of body contours was 67 HU.

Wang et al ⁶² adopted a conditional generative adversarial network (GAN) method, using coupled T1-weighted MRIs and CTs from 26 patients used for model training and 5 used for validation. The generated synCTs were evaluated by two radiologists, with excellent satisfaction in spatial geometry and noise level, good satisfaction in contrast and artifacts, and fair imaging details. The mean structural similarity indices between CTs and synCTs were 0.96 (cosine angle distance) and 0.84 (mean structural similarity).

Li et al ⁶³ adopted four different models using a GAN network and T2-weighted MRI and corresponding CT datasets of brain cancer patients from two different hospitals (one used as source and on as target): the best performance was achieved with a combined model using both source and target datasets, obtaining a MAE of 74.56 HU.

Koike et al ⁶⁴ used 580 pairs of CT and MR images to train a GAN algorithm to generate synCTs from MRI of 15 GBM patients: mean MAEs for the whole body, soft tissue and bone region were 108.1 ± 24.0, 38.9 ± 10.7 and 366.2 ± 62.0 HU. Treatment plans were generated with 3D-CRT and VMAT techniques on the synCTs (prescription dose of 20 Gy with or without a boost up to 40-60 Gy, 2 Gy/fr) and compared with those planned on CT: D2%, D50% and D98% for PTV1, and D_max for the brainstem, optic chiasm and left optic nerve were significantly higher in plans from synCTs, but differences were small and clinically negligible (<0.7%); gamma pass rates were 99.7% at 3%/3 mm, 98.9% at 2%/2 mm and 94.2% at 1%/1 mm.

Tang et al ⁶⁵ adopted a GAN using 27 pairs of MRI and CTs from brain cancer patients for training to develop synCTs from T1-weighted MRI of 10 other patients. Comparing CTs and synCTs, mean MAE was 60.52; the average pass rates of gamma analysis using the 3%/3 mm and 2%/2 mm criteria were 99.76% and 97.25%. No significant differences of dose-volume histogram were found between plans generated on CTs and synCTs for both target and OARs (lens, brainstem, optic nerves, chiasma, brain) with slight dose differences (range from −0.77% to 1.33%).

A deep learning architecture was implemented by Gupta et al ⁴⁸ to produce synCTs from T1-weighted MRI of 60 patients with intracranial tumors (47 used for training and 13 for test): MAE was 17.6 HU in soft tissue and mean geometric shift between CTs and synCTs was −0.1/-0.2 mm in all the axes, but a mean target dose difference of 2.3% was reported (mainly due to the lack of immobilization mask on the synCT images).

Kazemifar et al ⁴⁹ designed a GAN method to create synCTs from T1 -weighted spin echo MRI sets of 77 heterogeneous brain cancer cases (tumor volume 1.1-42.4 cm³) and compared 14 treatment plans generated on synCT and simulation CT. MAE was 47.2 HU and Dice similarity coefficient (DSC) was 80% in bone. Mean difference between doses calculated on synCT and simulation CT for all DVH parameters of PTV and OARs (brainstem, eyes, optic nerves, chiasm) was statistically insignificant and <1%; 3D gamma analysis reported mean passing rates of 99.2% and 94.6% using 2%/2 mm and 1%/1 mm acceptance.

Emami et al ⁵⁰ developed and validated a GAN deep learning model that generated robust synCT in seconds from a single T1-weighted MRI and outperformed a DCNN algorithm, with better representation of regions of abnormal anatomy.

The same group adopted this model to generate SynCT for 12 brain cancer patients (6 treated with conventional treatment and 6 with SRS, 8 adjuvant treatments and 4 ablative/radical all treated with VMAT) from T1-weighted postgadolinium MRI. ⁶⁶ Excellent dosimetric agreement was reported between simulation CT and synCT, with mean difference ≤0.10 ± 0.04 Gy for the target D95% and ≤0.13 ± 0.04 Gy for OARs; agreement in dose planes at the isocenter was optimal (mean gamma passing rates 99.9% and 99.0% at 2%/2 mm and 1%/1 mm, respectively). This algorithm allowed as well to generate adequate imaging for IGRT, with mean difference in CBCT-synCT registrations <0.2 mm and mean difference between kV-synCT DRR and kV-simCT DRR registrations <0.5 mm.

Dosimetric Feasibility of MRI-Linac Treatment Planning for Brain Metastases and Comparison With Conventional Plans

Treatment planning and delivery on MRL entails multiple challenges, such as the evaluation of dose distribution in the presence of a magnetic field and the necessity to calculate an adaptive plan online in limited times. Multiple sophisticated algorithms have been developed to take into account the transport of charged particles in a magnetic field and the influence of the Lorentz force on the secondary electrons, such as the stochastic Monte Carlo method or a deterministic grid-based Boltzmann solver^72–75 and satisfactory dosimetric accuracy was reported between original, adapted and delivered plans using dedicated phantoms. ⁷⁶

Several previous studies analyzed the feasibility of treatment planning for metastatic brain lesions in term of OARs sparing and target coverage, mainly simulating stereotactic single-fraction radiosurgery or hypofractionated radiotherapy and generally comparing the results with plans generated or delivered with conventional linacs.

The technical feasibility and systematic accuracy of MRIdian Linac for treating multiple (2-3) BM with a single isocenter was investigated by Wen et al. ⁷⁷ Five previously treated patients (11 lesions) were evaluated for dosimetric accuracy, the plans adopted 10 to 15 beams step-and-shoot IMRT with a prescribed dose of 16 or 18 Gy for SRS; a three-compartment phantom was used for the end-to-end test. Excellent plan quality and deliver accuracy could be obtained: end-to-end localization accuracy was 1.0 ± 0.1 mm and average absolute point dose difference between measured and calculated dose was 1.64%.

A comparison between plans simulated for MRIdian and VMAT plans was performed by Slagowski et al. ⁷⁸ Clinically acceptable plans were generated for single-isocenter brain SRS (step-and-shoot IMRT with 11 coplanar beams, 20 Gy single fraction) for MRIdian on the basis of data from six cases of BM previously treated on a conventional linac for lesions with a diameter ≤2.25 cm, while for larger lesions the normal brain V12Gy planning objective (≤10.0 cm3) could not be respected and was achieved only with conformal arcs and non-coplanar beams on the conventional linac.

Tseng et al ⁷⁹ retrospectively simulated three different plans for each patient with the same prescribed dose using 2-arcs non-coplanar VMAT for conventional linac and 9 beams coplanar step-and-shoot IMRT for conventional linac or Unity MRL both in the absence or presence of the 1.5 T transverse magnetic field. Data from 24 patients treated with VMAT for intact single BM, with SRS at a dose of 18 to 21 Gy (62.5%) or SRT at 24 Gy in 3 fractions (37.5%) were used. Although VMAT plans resulted in better dosimetric parameters, all the IMRT plans for MRL satisfied the target coverage and organs-at-risk objectives. The presence of the magnetic field determined a significant but minor increase in mean dose and D2cc to the skin (0.08 and 0.6 Gy, respectively) and around air cavities (0.07 and 0.3 Gy, respectively) that did not compromise target conformity or dose gradient.

In a recent dosimetric study, ⁸⁰ adaptive plans were generated using a dedicated phantom mimicking multiple brain lesions (prescription dose 20 Gy single fraction) or MRI from a patient treated with VMAT for two right frontal lesions (68 and 1.7 cm³, respectively) at a dose of 30 Gy in 5 fractions. Six preset and six random setup variations were simulated and two adaptive plans (step-and-shoot IMRT plans with 13 coplanar beams) per daily MR image were created for Unity using the ATP and ATS workflows; all adaptive plans were compared with the reference plan. Dosimetric goals were consistently met in both workflows, with the ATS workflow resulting in a reduction of brain V12 (−16/17%) and GI (−5/10%) and an increase in CI (+8%) compared with ATP in phantom models; GI, CI and V20 were as well better for ATS in the patient model and in both models and ATS plans were more consistent (±2-4% for the considered parameters) with reference plan compared with ATP plans; optimization time increased with ATS, but within a clinically feasible range (<60 s vs about 300 s). Difference between the planned and measured target dose were within 1% for both workflows.

A recent retrospective in silico study ⁸¹ adopted the data of 13 consecutive patients with resected single BM to simulate step-and-shoot IMRT SRS plans for Unity MRL directly on post-operative MRIs acquired a median of 2 days after surgery. These plans were compared with non-coplanar VMAT-based SRS plans simulated on post-operative MRI and the clinically delivered non-coplanar VMAT-based SRS (prescribed dose 15-21 Gy, depending on the cavity volume) with CT-based IGRT contoured on post-recovery MRI performed a median of 20 days after surgery. Contours were performed on a 1.5 T MRI acquired containing a 3D T1 spoiled gradient echo (T1-TFE) gadolinium-enhanced scan. Target coverage and OAR constraints were met for all plans; median CI was equivalent (0.9) for all plans and IMRT plans had higher median gradient index (3.6 vs 2.7). Although volume of brain-GTV receiving 2, 5, 12 and 14 Gy and V3Gy of the skin (18.4 vs 1.1 cc) were significantly higher for post-operative IMRT plans compared with post-operative VMAT plans, all the plans were clinically acceptable.

Treatment plans for whole brain radiotherapy with hippocampal avoidance (HA-WBRT) were simulated with step-and-shoot IMRT for MRIdian system from the data of 12 patients treated on a conventional linac with VMAT. ⁸² All the MRgRT plans met the dosimetric constraints and were clinically deliverable, with comparable quality and delivery accuracy compared with VMAT plans, although significantly less homogeneous and with acceptable but significantly higher average hippocampi D100% (8.62 vs 7.92 Gy) and non-significantly higher maximum dose to hippocampi (15.00 vs 14.19 Gy) and optic structures (31.26 vs 30.94 Gy) and longer delivery times.

The administration of contrast medium before simulation MRI and, particularly in case of SRS or SRT, daily adaptive plans would improve the definition of tumor volume.

The dosimetric impact of gadolinium contrast medium for brain cancer in the presence of a magnetic field has been evaluated by Fujita et al: when the GTV in a brain tumor was overwritten with Gadovist, mean dose incremented by 8.90% in the absence of a magnetic field, by 4.81% in the presence of 0.35 T, by 1.76% in a 1.5 T field and 0.65% in a 3 T field. ⁸²

Ahmad et al evaluated the impact of the presence of various concentrations (8-157 mg/mL) of gadolinium-based contrast medium for the treatment planning on Unity MRL using a phantom model or data from 6 GBM patients (adopting 7 fields IMRT). Dosimetric differences could be quantified in Monaco for concentrations higher than 23 mg/mL (dose difference of D50 in the GTV of 3% at 1.5 T), while for lower concentrations a correction might not be necessary. ⁸³

Therefore, as the estimated concentration of Gadolinium contrast media in brain tumors is much lower (0.00073 mmol/mL, equivalent to 0.44 mg/mL) ⁸⁴ its dosimetric impact should not be clinically relevant.

Dosimetric Feasibility of MRI-Linac Treatment Planning for Primary Brain Tumors, Comparison With Conventional Plans and Clinical Feasibility

A few published experiences reported the results of treatment plan calculation for primary brain cancers on MRL and preliminary clinical experience.

Using data from 6 GBM patients treated with VMAT on a conventional linac at a dose of 60 Gy/30 fr, Ruschin et al ⁸⁵ quantified the dosimetric impact of adopting virtual couch shift correction for 2 to 6 mm translational errors on plans calculated for Unity MRL with methods based on shift-only, segment weight (SWO) or segment weight and shape (SSO). All the plans were clinically acceptable, with only a slight (1.2-2.0%) increase of brainstem D0.01cc in adapted plans and a reduction with SWO and SSO in the D0.01cc to other OARs (eyes, lenses, optic chiasm and optic nerves) and in brainstem D50; CI remained unchanged between reference plans regardless of correction method. Magnetic field had no clinically relevant effect on relative variations between reference and corrected plans.

Ding et al ⁸⁶ extracted the data from 6 patients with primary brain tumors (dose 60 Gy in 30 fractions) treated with a conventional linac and generated new plans with Monaco planning system for dynamic multi-leaf collimator (dMLC) IMRT or step-and-shoot IMRT for a conventional linac and a comparison was performed with plans generated with step-and-shoot IMRT (9 fields) for Unity MRL. All the plans calculated for MRL were clinically equivalent to those adopted for clinical practice. Compared with IMRT plans for conventional linac, MRL plans had acceptable but statistically significant increase in point maximum dose in the lens and worse homogeneity index, slight increase in optic chiasm and optic nerve D_max and similar mean dose to the optic chiasm, optic nerve and eyes. Calculation of MRL IMRT plans required less optimization time (42.3-64.2% reduction) and increased the number of monitor units.

A recent prospective cohort study ⁸⁷ evaluated 37 glioma patients (64.9% GBM) enrolled in the MOMENTUM trial that received at least one fraction of adjuvant RT on both Unity MRL and on a conventional linac; MRL plans (≥9 coplanar fields step-and-shoot IMRT) were calculated with a Monte Carlo-based algorithm and plans for conventional linac (single-energy 6 MV, ≥ 7 coplanar fields step-and-shoot IMRT with an additional non-coplanar field if needed) with a convolution-based algorithm. Monaco could generate safe plans achieving all the planning objectives, but with more heterogeneous dose distributions (significant increase in PTV D50%, D5%, and D2%) and slight but significant increase in brainstem D0.1cc, eye globe and lens D0.03cc, while no significant differences were detected for D0.03cc of other OARs (optic chiasm, optic nerves, cochleae); on the other hand, dose conformity was higher and falloff outside the target was equivalent.

It has to be noted that MRL plans resulted in 1.52 Gy higher Dmean (P < .0001) and 1.23 Gy higher D2cc (P = .0007) to tissues surrounding air cavities, while skin Dmean was 1.10 Gy higher (P < .0001), and skin V20Gy was 19.04 cm³ larger (P = .0001) compared with plans for standard linac. In vivo skin dose measured in 10 patients with an optically stimulated luminescent dosimeter was 14.5% greater for MRL plans (P = .0027) and was more accurately predicted by Monte Carlo-based calculation.

Contouring recommendations have been provided by a consensus from the MR-Linac International Consortium Research Group, ⁸⁸ demonstrating a high level of agreement in a MRI-only workflow among 6 experts, with no significant differences with the CT-MRI workflow for the contouring of 10 heterogeneous cases (5 glioblastoma and 5 grade 2 or 3 gliomas) on the basis of post-gadolinium T1-weighted MRI and T2/FLAIR MRI.

In order to optimize patient set-up, specific MRI-safe immobilization devices have been developed including commercial and in-house solutions, taking into account the inclusion of the coils and hearing protection and in order to offer an adequate comfort during the prolonged treatment time required by MRL.^89,90

Tseng et al ⁹¹ reported the initial experience of ten high grade glioma patients (4 GBM) that completed chemoradiotherapy on Unity MRL at a median dose of 60 Gy in 30 fractions. Online plans were generated according to the ATP workflow on a T1-weighted sequence, IMRT optimization typically involved 9 beams and mean in-room time per fraction for treatment was 37.3 min. No grade ≥ 3 acute toxicities were reported and only 2 patients experienced grade 2 side effects; despite possible concerns of claustrophobia, none of the patients discontinued the treatment. Acquisition of multi-parametric images (including DWI, CEST and blood oxygenation level dependent - BOLD) was feasible during routine treatment. At 10, 20 and 30 days after radiotherapy start respectively 3, 3 and 4 patients had a FLAIR volume that changed by at least 20% and three patients required replanning (between fractions 15 and 19) due to increase of target lesion or edema volume. In the four patients that progressed during follow up, region of recurrence included most of the low-ADC region identified during radiotherapy, with a positive predicting fraction >68% by the final fraction.

A retrospective study by Kozak et al ⁹² evaluated the clinical application of Unity MRL to treat pediatric and adolescent patients, including two patients with recurrent diffuse intrinsic pontine glioma: two male adolescents (one 13 years and one 19 years old) received re-irradiation at a dose of 31.2 Gy in 26 BID fractions and treatment was well tolerated

Assessment of Target Volume Changes Before and During Treatment

Primary Brain Tumors

Current guidelines suggest the acquisition of a dedicate planning MRI to delineate contours for adjuvant RT for GBM in addition to post-surgical MRI, that is usually performed within 72 hours after surgery. ⁹³ Indeed, several studies performed in patients that underwent adjuvant RT on conventional linacs demonstrated relevant changes of target volume between post-operative MRIs and planning MRIs^94,95 or MRIs performed at the first day of RT^96–98 and this could affect CTV definition. ⁹⁴ Moreover, repeated MRIs performed during RT course invariably resulted in volumetric and/or geometric variations of disease and target volumes^98–102 and allowed to improve doses to the OARs^99,100,102 or to improve target coverage.^99,102 Volumetric variation of target volumes was heterogeneous, but in most series the majority of lesions progressively shrinked during radiotherapy^98–100 although expansion was consistently reported in a fraction of the cases.^{98,99,101,102} Adaptive radiotherapy based on serial MRI findings could improve doses to the OARs particularly in case of volume reduction^99,101,103 and guarantee adequate coverage in case of volumetric increase. ¹⁰³ An example of volumetric changes during radiotherapy in a glioblastoma patient are represented in Figure 2.

Figure 2.

Imaging of a patient treated with adjuvant radiotherapy after radical resection of glioblastoma on Unity MRL with concurrent temozolomide (dose 60 Gy in 30 fractions). A, simulation T2 MRI performed on diagnostic 1.5 T scanner; B, simulation CT scan; C, T2 MRI performed with the on-board 1.5 T scanner at first radiotherapy fraction; D, – T2 MRI performed with the on-board 1.5 T scanner at tenth radiotherapy fraction; E, T2 MRI performed with the on-board 1.5 T scanner at eighteenth radiotherapy fraction; F, T2 MRI performed with the on-board 1.5 T scanner at last radiotherapy fraction.

However, in real-world clinical practice the limited availability of MRI does generally not allow to perform dedicated simulation MRI and, particularly, repeated MRI during treatment course. ¹⁰⁴ On the other hand, the current standard represented by simulation CT scan and CT-based IGRT is not sufficient to detect most of the tissue changes during treatment. The adoption of MRL could overcome these limits, allowing daily replanning based on the actual anatomy.

A preliminary analysis by Mehta et al ¹⁰⁵ evaluated volume changing of cavity, edema, and visible tumor during radiotherapy in three patients receiving adjuvant tri-cobalt-60 MRIgRT on MRIdian with concurrent temozolomide (60Gy/30fr) after surgery for GBM. Daily scans were performed with balanced steady state free precession sequence (a blend of T1 and T2, with dominant T2 weighting) showing a general trend towards volumetric reduction of the cavity, with a stabilization around fraction 20.

The analysis of daily MRI performed in patients with high grade glioma undergoing radiotherapy on Unity MRL in the context of MOMENTUM study ³² demonstrated detectable GTV changes for 14 of 20 cases (4 increased, 7 decreased, 3 both increased and decreased) with changes in median ADC of −0.5%, −0.5%, and 2.4% at 1, 3, and 6 weeks.

In the above-mentioned experience by Tseng et al ⁷⁸ including ten high grade glioma patients (4 GBM) that completed chemoradiotherapy on Unity MRL, FLAIR MRI performed at 10, 20 and 30 days after radiotherapy start revealed a volume change of at least 20% in 3, 3 and 4 patients respectively and 3 patients required replanning (between fractions 15 and 19) due to increase of tumor or edema volume.

Secondary Brain Tumors

BMs and surgical cavities undergo relevant volumetric and geometrical variations over time, as reported by multiple studies summarized in Table 2.^81,106–111

Table 2.

Tumor and Cavity Volume and Shape Variations of Cerebral Lesions Over Time.

Author	Patients/lesions	Interval between MRIs	Results
Seravalli et al 67	13 patients with resected single brain metastases	2 days vs 20 days after surgery	15.5% of cavities shrunk by > 2 cc, 46% expanded by ≥ 2cc; potential variation in dose prescription for 4 patients.
Hessen et al 93	42 intact brain metastases from 26 patients	Diagnostic vs simulation MRIs (median time 22 days)	Median center of mass shift of 1.3 mm (maximum shift of 5.0 mm) and a median distance between the tumor borders of 1.9 mm (maximum 7.4 mm).
Jarvis et al 94	43 resected brain metastases	Before surgery vs within 24 hours after surgery vs SRS planning	Between post-surgical and planning MRIs 46.5% of cavities stable in size, 23.3% collapsed and 30.2% increased by >2 cm3; 10 patients had cerebral disease progression (5 within the cavity, 4 elsewhere and 1 both).
Scharl et al 95	57 patients with resected single brain metastases	Post-operative vs planning contrast enhancing T1-weighted MRI (median interval 22 days)	Mean cavity size significantly decreased by 23.4%, volume reduction detected in 79.1% of cases and increase in 17.4%; suspect and confirmed local progression in 4 and 5 patients, respectively.
Kubo et al 96	23 patients with 27 intact brain lesions	MRI before RT vs during treatment (usually before or after the fourth fraction; 35 Gy in 5 fractions or 40 Gy in 8 fractions)	Plan modification in 55.6% of lesions (in 6 for tumor reduction, enlargement in 3, displacement in 3 and shape change in 3); change in PTV up to 43%.
Lee et al 97	40 intact brain metastases from 33 patients	MRI before the first fraction and after 1 or 2 fractions of SRT (performed in 3-5 fractions)	decreased tumor volume for 45% of lesions.
Hessen et al 98	18 lesions treated with SRT only, 20 lesions treated with post-operative SRT	MRIs performed at treatment planning vs during treatment (after 1-3 fractions, median interval 9 days)	median maximum distance between contours was 2.1 mm for SRT only and 2.0 mm for adjuvant SRT; PTV of lesions treated with SRS only significantly increased (median +0.7 cm3) with a decline of PTV dose coverage up to 34% (median 3.2%)

Cerebral metastases and peri-tumoral edema are subject to changes between diagnostic and simulation imaging, with relevant tumor shift and volume changes, also secondary to steroid administration. ¹⁰⁶ Moreover, due to the limited availability of MRI examinations, in clinical practice simulation MRI is often not performed and treatment plan relies on fusion with diagnostic imaging.

Surgical cavities after metastasectomy are as well subject to variation in volume and shape that can occur over a few days^107,108 and, although MRI is generally routinely performed after BM resection, post-operative imaging modalities and timing are characterized by high heterogeneity. ¹¹² Delaying SRS to surgical cavity of BM could negatively impact local control of disease ¹⁰⁷ and shortening the interval between surgery and radiotherapy could as well improve overall survival. ¹¹³

Moreover, brain lesions are subject to volumetric and geometric modifications during RT, even in case of extremely hypofractionated schedules.^109,110 Steroid administration and histology are predictive factors for change during RT ¹⁰⁹ and variations before treatment (eg between diagnostic and simulation MRI) could help to predict which lesions are more subject to a worsening of dose coverage during RT. ¹¹⁰

The adoption of MRL for the treatment of BM could overcome the above mentioned limits, allowing treatment planning on the actual anatomy observed on MRI and adaptive treatment in case of relevant variations.

Tan et al ¹¹⁴ evaluated CTV changes on 15 patients treated with Unity MRL for resected BM (27.5 or 30 Gy in five fraction) using an ATP workflow, gadolinium-enhanced T1 (T1c) MRI was acquired at planning (median 15 days after surgery) and fraction 3 and T2/FLAIR sequence at simulation and at every fraction. A reduction of T1c CTV volume was detected for 80% of patients at fraction 3 with a significant volume reduction from 12.0 cm3 and 10.2 cm3 to 9.3 cm3 and 8.6 cm3 (- 11.4% and - 8.4%) in T1c and T2/FLAIR, respectively; in the other 20% of the cases CTV increased (median change +7%). Additionally, significant variability in CTV contouring was observed between T1c and TS/FLAIR MRI, with larger contours in T1c in 80% of the cases (median 6% volume increase) suggesting the use of T1c for contouring. The effect of CTV contraction translated into increased dose to healthy brain in plans retrospectively generated by re-contouring the CTV on T1c performed at fraction 3 using the original and unaltered reference plan: there was a significant increase of V30Gy and V25Gy of healthy brain; two of 12 patients with a CTV reduction developed asymptomatic radionecrosis on follow up MRI.

Conclusions

The adoption of an online MR-guided workflow could offer several advantages for the treatment of brain tumors, including the minimization of image registration error, superior image quality for intra-cranial lesions and OARs and the possibility to adapt the treatment plan on the basis of daily MRI. On the other hand, multiple issues must be addressed in order to implement the shift from a CT-based to an MRI-based workflow.

To the best of our knowledge, only one review focusing on the use of MR-Linac for the treatment of brain cancers in 2017, ¹⁹ therefore the majority of the available literature was not yet published.

Specific correction algorithms and dedicated solutions allowed to reduce image distortion well below 1 mm, with residual geometric discrepancies at air-tissue interfaces. Multiple sequences have been optimized in order to acquire images with adequate quality with on-board MR scanner in limited times.

To allow dose calculation on MRI, electron density must be assigned through the generation of ‘syntetic-CT’, that can be implemented through several strategies (including bulk density assignment, atlas-based methods and, particularly, machine learning techniques) with excellent geometric accuracy and sources of uncertainty represented by abnormal anatomy due to large lesions, difference in set-up, air cavities and variability in bone density among individual patients. Plans generated on syn-CTs demonstrated optimal dosimetric agreement with those obtained on CT scans.

Various sophisticated algorithms have been developed to compensate the impact of magnetic field on dose distribution and calculate daily adaptive plans in limited times.

Previous dosimetric studies demonstrated the feasibility of delivering SRT for BM on MRL in terms of OARs sparing and target coverage, with excellent plan quality and delivery accuracy. Clinically acceptable plans, with adequate target coverage and respect of OARs constraints were consistently generated. Similar results were reported in the few published experiences assessing treatment plans generation for primary brain tumors on MRL.

It has to be noted that VMAT plans for conventional linacs generally resulted in better OARs sparing and significant but minor and clinically acceptable increase in dose to the skin and around air cavities have been reported. It should as well be considered that VMAT technique is currently being implemented on Unity system and should be clinically available in next years. Therefore, as there are still no clinical data regarding the trade-off between the adoption of more conformal techniques on ‘conventional’ linacs and the benefit provided by daily replanning with improved image quality on MRL, treatment should be personalized and take into account disease and patient characteristics (eg close proximity to critical OARs, potential tumor variation during treatment).

Patient selection could therefore be crucial to define not only patients suitable for treatment with MRgRT (eg contraindications such as claustrophobia or non-MR safe implants) but as well subjects that would benefit more from the unique features offered by MRI-Linacs.

Nonetheless, the main advantage of treating brain tumors with a MRL is represented by the possibility to detect target volume changes that occur between diagnostic imaging and treatment planning and as well during treatment and perform a daily adaptation of treatment plan.

Indeed, multiple studies reported major variation in target volume and shape between post-operative and planning MRI and between MRIs performed at different times during radiotherapy for primary brain tumors. These changes were case-specific, generally represented by shrinkage but encompassing as well enlargement in a relevant fraction of patients, leading to excessive doses to the organs at risk and/or undercoverage of PTV.

BMs are as well subject to geometric changes and position shift even over limited time intervals, secondary to tumor growth and variations in peritumoral edema and dependent on tumor histology and steroids administration. Surgical cavities after metastasectomy are likewise dynamic in size and shapes, with heterogenous changes.

Adaptive radiotherapy based on daily MRI could improve OARs sparing, particularly in case of volume reduction, and guarantee adequate coverage in case of volumetric increase.

Another promising field of study is represented by the opportunity to perform functional MRI studies, that can be practicably integrated in the clinical workflow and could provide markers linked with toxicity and/or response to treatment.

Multiple issues should as well be considered for the selection of the patients that might benefit more from a treatment on MRL. Indeed, although currently only a few papers evaluated the tolerability of the treatment for patients with brain tumors with no feasibility concerns, the relatively long treatment time with an immobilization mask (that might exceed 40 minutes) could result in excessive discomfort for some subjects. For this reason, in the context of clinical practice and outside of the scope of clinical trials the number of performed sequences should be limited in order to avoid prolongation of the time on the table.

Intra-fraction motion should as well be considered, although it is limited for intracranial sites and can be controlled with real-time imaging during the session, as both the clinically available systems offer the possibility to acquire real-time imaging during treatment at 4 to 5 FPS.

Another limit of the diffusion of MR-Linac systems is the increased capital required for installation compared with a ‘conventional’ Linac and the slow throughput due to long treatment times limits daily capacity; nonetheless, economic analyses demonstrated that the use of MRgRT for selectet complex cases that benefit form adaptive treatment might be as well cost efficient. ¹¹⁵

Preliminary clinical experiences reported fair tolerability and optimal toxicity for patients treated on MRL for primary brain cancer and the increasing number of facilities implementing MRL will lead to a growing body of literature that will clarify the potential benefit for the treatment of brain cancers.