Approximately half of patients diagnosed with cancer will be treated with radiation therapy (RT)1 Treatment efficacy depends on accurate and precise delivery of radiation to the target area, while minimizing radiation to surrounding structures or organs at risk (OARs). Technological advances have allowed conformal delivery of radiation dose with intensity-modulated radiation therapy (IMRT), volumetric modulated arc therapy (VMAT), and stereotactic body radiosurgery (SBRT)2 improving the therapeutic ratio.

Highly conformal delivery techniques with steep dose gradients necessitate high-quality image guidance to ensure targeting accuracy. Image guided radiation therapy has evolved considerably since the 1990s, particularly with respect to routine use of cone-beam computed tomography (CBCT)3 However, CT imaging has notable limitations including suboptimal soft tissue contrast (fiducial markers are routinely used as a surrogate of the target position to compensate for this) and an inability to acquire images during treatment delivery. Moreover, conventional CT-guided linear accelerators have so far been unable to perform rapid online adaptive replanning to account for interfraction anatomic changes.

Magnetic resonance imaging (MRI) linear accelerators (LINACs) pair both technologies to provide increased soft tissue contrast in real time. MRI-guided radiation therapy (MRIgRT) utilization has increased exponentially over time, especially for the treatment of pancreas, prostate, lung, and liver malignancies.4,5 In this review, we discuss the advantages of MRIgRT and the process of online adaptive treatment planning and quality assurance.

There are numerous advantages to MRIgRT including superior soft tissue contrast compared to CT, continuous intrafraction MR imaging, and online adaptive replanning. These technical advantages translate into important clinical benefits including the ability to safely dose escalate while reducing toxicity and reducing fractionation. As of 2020, approximately 70% of MRIgRT treatment courses were ultra-hypofractionated (1-5 fractions).5 Online adaptive MRIgRT that is done while the patient remains in the treatment position incorporates real-time anatomical changes such soft-tissue deformity, volume changes, and changes in OAR positioning. Improved delineation of internal structures enables decreased planning target volume (PTV) margins and therefore OAR doses.

Intrafraction MR imaging of internal anatomy decreases target and OAR uncertainty and enables detection of volume changes due to treatment. Multiple studies have shown that MRIgRT can reduce PTV sizes. For example, during prostate SBRT, high dose regions can overlap with the bladder and rectum, causing toxicity. Reducing the PTV margin from 4 mm with CBCT versus 2 mm using MRIgRT decreases grade 2+ genitourinary toxicity by an absolute difference of 19% and gastrointestinal (GI) grade 2+ toxicity by 10.5%.6 Similarly, in patients with local recurrence after prostatectomy, their PTV expansion was isotropically reduced from 5 mm to 3 mm on CT-guided and MRIgRT RT, respectively. MRIgRT was associated with a 30.5% reduction in any grade acute GI toxicity.7

For structures that are prone to movement, continuous cine-MR imaging during treatment enables advanced motion management and enables real-time position tracking. For example, the prostate is known to move during treatment and swell during a course of SBRT.8,9 Intrafraction cine-MRI monitors the prostate position during dose delivery. An automatic beam-hold can be initiated if the prostate moves beyond a defined gating boundary.

For example, in the treatment of localized prostate cancer and prostate cancer recurrence using a 0.35 T MRI-LINAC, a 3 mm gating boundary was placed around the prostate and prostatectomy bed. If 10% of the PTV moved outside of the gating boundary, the radiation beam was held.6,7 Advanced gating is particularly useful in the treatment of peripheral lung cancers, which can move more than 2 cm during respiration.10 Clinical studies have shown that gating breath-hold delivery can result in a reduction by approximately half of the treatment volume compared to the internal target volume (ITV) plus a standard 5 mm expansion when using free-breathing 4D-CT scans.11 Similar approaches are also feasible with primary and secondary hepatic neoplasms. For example, combining gating protocols with a deep inspiration breath hold (DIBH) has led to a freedom from local progression of 90% to 100% and 75% for hepatocellular carcinomas and colorectal metastasis, respectively.12,13

Incorporation of MRIgRT-based functional imaging has the potential of detecting biochemical changes that precede anatomical changes.14,15 Functional imaging can detect changes in tissue perfusion and oxygenation, which can be correlated to early treatment response or resistance.16 As a proof of concept, MRI-LINAC mDixon, T1-weighted, and T2-weighed sequences have been obtained in healthy volunteers.17 Chemical exchange saturation transfer (CEST) MRI can acquire metabolic information that has been shown to predict therapeutic response in glioblastoma (GBM), distinguish radionecrosis from tumor, and differentiate pseudoprogression from true progression.18, 19, 20, 21 CEST imaging is possible using a 1.5T MRI-Linac.22 Characterization of changes during treatment and on-treatment adaptation to encompass areas of true progression and spare areas of radionecrosis may improve outcomes in patients with CNS tumors.

Online treatment adaptation has the potential of incorporating interfraction volume changes to improve dosimetric parameters. Online adaptation is not MRI-specific and has been utilized using daily CBCT images in the treatment of head and neck cancers, adapting to changes in neck diameter due to edema or weight loss and shifts in bony structures and target volumes.23,24 Using adaptive CBCT-guided radiation therapy approximately 10% to 30% of patients have required re-planning, especially when treated with chemoradiation.24,25 Adaptation reduced OAR doses including spinal cord, brainstem and parotid glands, while boost planning volume to the primary site improved by 1%.26

Poor CBCT soft-tissue contrast can hinder the adaptive workflow. Single institution pilot studies have shown clinical feasibility of using daily adaptive MRIgRT with minimal differences in reference and summation plan doses. To date, most adaptive planning has been performed to minimize OAR dosing.27 There are currently multiple phase II clinicals exploring the use of dose de-escalation in HPV+ oropharyngeal cancers and OAR sparing in head and neck cancers.28 (NCT03224000). Future studies will need to be conducted to determine if MRIgRT is superior to CBCT-guided radiation therapy for adaptive head and neck cancer treatments.

Treatment adaptation varies by cancer type and location. For example, studies show that approximately 10% to 12% of localized prostate cancer require adaptation, while over 90% of locally advanced pancreatic cancer treatments require adaptation.6,29 Online adaptive MRIgRT for peripheral lung cancers marginally benefitted from re-adaptation when gating and real-time MRIgRT were employed.11 In contrast, central and ultracentral tumors were seen to benefit more from adaptation due to their proximity to more radiosensitive structures18; more than 90% of fractions were replanned. Replanning decreased the percent of violated dosimetric constraints from 94% to 17%.18 Future studies will need to determine if treatment adaptation is efficacious, especially in cancers that are less likely to require adaptation.

Dose escalation has been hypothesized to improve overall survival in some patients, for example, those with locally advanced pancreatic cancer.30,31 However, higher doses are limited by small bowel dosimetric constraints.32 A Dutch phase II trial used a 10 mm CTV margin and a prescription of 45 Gy in 3 fractions using CT-based SBRT approaches. Approximately 94% of their patients developed at least grade 2 pain and many patients developed severe GI ulceration or perforation.33 MRIgRT improves pancreatic cancer visualization, allowing its position relative to radiosensitive structures to be determined. Using the 0.35 T MRI-LINAC system, mid-respiratory breath hold, automatic gating, and adaptive planning, allowed the margin to be reduced by 3 to 5 mm. Using these approaches, there were minimal acute and late toxicities of 2.9% and 2.9%, respectively, while grade 3 toxicities were uncommon. Local control and overall survival were also improved relative to historical controls.29 Of note, approximately 96% of the fractions were adapted to daily positional changes. These data indicate that treatment of pancreatic cancer could benefit greatly from MRIgRT-based adaptive planning.

In patients with oligometastatic disease, it might be possible to dose escalate to targets that are near radiosensitive structures. Use of stereotactic online-adaptive MRIgRT (termed SMART) was evaluated in patients with oligometastatic or unresectable primary abdominal malignancies.34 Approximately 83% of fractions required online adaptive plans due to predicted OAR constraint violations or the opportunity to dose escalate. No grade 3 or greater toxicities were reported. Numerous clinical trials are ongoing to evaluate the use of MRIgRT to treat oligometastic disease.35 Single-fraction ablative radiotherapy in patients with primary and metastatic disease is under investigation in the SMART ONE trial (NCT04939246). Other studies are looking at MR-only simulation (no CT simulation) treatment planning in patients with oligometastatic disease, which would allow patients receive their first fraction the same day as their MR simulation (NCT03878485, NTC03824366). CT-simulation-free planning and same-day treatments could potentially expedite patient care.

Currently 2 MRI-LINAC systems are clinically available, a 0.35 T (MRIdian, ViewRay, Inc, Mountain View, CA) and a 1.5 T (Unity, Elekta AB; Stockholm, Sweden) system. Different design tradeoffs were made to overcome engineering challenges to integrate a LINAC and an MRI.36 The magnetic fields interact with the components of the LINACs, such as the waveguides, radiofrequency sources and the multileaf collimator motors, and therefore shielding is needed. Also, MR scanners rely on sensing very weak radiofrequency signals that could be contaminated by the radiofrequency source and various motors of the LINAC, introducing artificial noise. Rotating metal parts of the LINAC or its accessories could also perturb the B0 field homogeneity. Detailed information about the technical concepts can be found in the literature.37,38 Owing to the perpendicular orientation of the magnetic field with respect to the radiation beam, which influences the dose distribution, both systems use dedicated treatment planning systems (TPS) with Monte-Carlo-Dose calculation.39

Since both systems combine closed MRI scanners with LINACs, the adaptive patient treatment workflows are similar in most aspects, as schematically presented in Figure 1.40,41 The first step in the clinical workflow is a simulation session that includes a simulation CT and simulation MRI in the treatment position with the dedicated MR-safe immobilization devices as per the clinical site-specific protocols. The planning MRI can either be conducted on the MRI-LINAC itself or on a separate MR scanner for radiotherapy simulation. In both cases, dedicated radio-transparent receiver coils are used to prevent deformation of the patient outline while minimally attenuating the radiation beams. Based on these simulation images, a treatment plan, also called a reference plan, is optimized and clinically approved by the physician and physicist. Depending on departmental procedures, IMRT dose distributions quality assurance may be required by irradiating dedicated phantoms using the treatment plans comparing the results against calculated dose distributions.42, 43, 44, 45

Each treatment session starts with a daily MRI to validate patient positioning and decide on an adaptation strategy. For a non-adaptive workflow, the daily MRI is registered rigidly to the reference simulation image and an isocenter shift is established. For the 0.35 T MRI-LINAC the isocenter shift is compensated for by a 3-degree of freedom couch translation. The couch of the 1.5 T MRI-LINAC cannot translate, so a virtual couch shift is used. The beam segments are shifted accordingly and the leaf sequences modified to reflect the flattening-filter free fluence distributions.40 In that case the reference image with updated isocenter position is used as frame of reference for dose calculation. As daily anatomical changes are not represented, violation of dose-volume criteria would only be apparent if the dose distribution is recalculated on the daily MRI.40 Therefore, this workflow should only be used if anatomical changes were assumed negligible.

For the online adaptive replanning workflows, also referred to as ‘adapt to shape (ATS)’, the OARs and targets are contoured on the daily MRI46 (Fig. 2). Using the deformation vector field of a deformable registration the contours of the reference image are propagated to the daily MRI. Afterwards contours can be manually adapted if necessary. To minimize contouring time, it has been proposed to re-contour only a volume within 3 cm expanded from the PTV surface.47 It should be considered that such strategies require a modification of treatment plan evaluations, as for instance dose-volume criteria based on relative volumes are no longer meaningful.47 Since this workflow uses the daily image for optimization and dose calculation, it accounts for anatomical changes, including changes in size and deformation, as well as the spatial relationship between the targets and OARs.

MRI scans do not readily provide the electron density distributions needed for dose calculations, so strategies for generating synthetic CT scans are needed. Electron density information can be obtained from the simulation CT using bulk density overrides of contoured structures or by deforming the bulk density map with the deformation vector field gained using deformable image registration. Pham et al.48 demonstrated that correct mapping of electron densities for air cavities is challenging, as there can be tremendous differences in size, shape and location between reference and daily image. Once the contours and electron density information are generated, a new treatment plan can be optimized. Based on the size of anatomical changes and to rationalize optimization time both commercial MRIgRT systems provide strategies with a range of complexities.40,41,49 Depending on the anatomical changes a so-called warm start optimization, which starts with the beam segment shapes and weights of the reference plan and adapts both or a full optimization from fluence can be chosen. To facilitate the process, many of the treatment plan parameters, such as beam angles, segmentation restrictions and planning constraints are copied from the reference plan but are thereafter alterable. After treatment plan optimization and plan approval, the plan is transferred to the LINAC and ready for radiation delivery.

Both systems offer fast planar-imaging, also called cine imaging, to monitor tumor position during the application of radiation dose. For automated radiation beam gating the tumor needs to be automatically tracked on the cine-images. Only when the tracked position overlaps with a preprogrammed gating window will the radiation beam be turned on. For the 0.35 T MRI-LINAC automatic real-time tumor tracking and beam gating has been available since 2017.50 The automated tracking and gating system for the 1.5 T MRI-LINAC was officially introduced in October 2022 and broad clinical implementation is expected in 2023.

Currently available hardware and clinical workflows for online adaptive MRIgRT as described above offer unique opportunities to combine the advantages of MRI with those of adaptive RT in a single treatment session. Clinical challenges implementing MRI-LINACs were associated with ensuring MRI patient safety as well as claustrophobia, and in some cases the limited radiation field size.51, 52, 53, 54 Dedicated analyses identified additional risks related to MRIgRT adaptive treatments relative to IGRT with conventional LINACs.41,55 Despite the numerous advantages provided by the innovative MRI-LINAC technical design, the clinical workflow is a time-consuming and fragmented process consisting of several distinct and often manual steps. These workflow steps include the acquisition of the daily MRI scans, registration of the daily images to the reference image, manual or semi-automated recontouring of relevant OARs and the CTV or GTV, creation of dependent volumes of interest such as planning risk or target volumes (PRVs/PTVs), reoptimization of the treatment plans, acquiring additional MRIs for repeat position verification, quality assurance, and finally application of the radiation dose including real-time cine imaging and gating (cf. Fig. 1). These workflows were not the responsibility of a single professional group, they included radiation oncologists, medical physicists, dosimetrists, and therapists. Initial reports on early experiences and operating procedures described increased staffing requirements.51,54 Furthermore, several groups reported that each workflow step took 5 to 15 min, leading to 30 to 60 min treatment sessions.51,54,56,57 This amount of time could have undesired consequences such as intra-fraction organ motion or volume changes which subsequently could negatively impact plan quality.58,59 To overcome the limitations associated to the relatively large treatment session length, several groups have proposed strategies such as dividing the daily treatment fractions into several sub-fractions or performing a second, fast adaptation consisting of an isocenter-shift or adapt-to-position based on a fast MRI scan for final position verification after time-consuming plan reoptimization just before treatment application.60, 61, 62

To gain the full capacity of MRIgRT in the future, real-time application of MRI-guided adaptive radiotherapy will be needed. This will require the automation of each workflow step with extremely short latency times, including quality assurance and surveillance mechanisms to ensure patient treatment safety and accuracy. Several investigators have shown the potential of using deep learning strategies for adaptive MRIgRT by either training AI models for MRI-based fast automatic organ-at-risk and target contouring or robust propagation of structures from the reference image to the anatomy of the day, including prediction of moving structures.60,63,64 Robust and accurate contouring is essential because contouring errors can cause substantial dosimetric errors.65 Furthermore, different fully automated treatment planning solutions have been recently proposed which could speed up the online MRIgRT planning step.66, 67, 68 Other groups have demonstrated that automating the treatment planning and optimization process could help the decision making process related to the need for daily adaptation.69,70 Deep learning approaches could open up novel strategies to perform automatic treatment planning and dose calculations.71,72 Automatization of each workflow step in the online adaptive MRIgRT workflow has been shown to be feasible with clinically acceptable quality, and will thus speed up the process to potentially real-time in the near future.67,68,73,74

To ensure patient safety during online adaptive MRIgRT, integrating quality assurance (QA) and radiation safety checks into the online workflow is critical. Therefore, time-efficient QA procedures are required. Because adapted plans can significantly differ from the refence plan and no measurement-based QA can be conducted, fast, accurate and automated, online QA is needed. Different approaches, including secondary independent dose checks, plan sanity or plausibility checks, portal dosimetry, and log file analyses have been recently proposed.49,75, 76, 77, 78, 79, 80, 81, 82 The novelty of the adaptive workflow may also require new QA checks, eg, for contouring and synthetic CT generation.49 After treatment, dose distributions can be verified using dedicated phantoms and machine QA.42, 43, 44, 45,49,83,84 Most of the proposed tools have been in-house-developed solutions and to date, no consensus has been developed regarding which online quality checks should be required.

A further and crucial aspect in online adaptive MRIgRT is related to the accurate reporting of the tumor and OAR radiation doses. MRIgRT plans are adapted daily due to positional, anatomical and functional variations, so the OAR doses may vary widely and may require per-fraction deformable dose accumulation. Recent expert reviews have discussed practical considerations on using deformable dose accumulation, also called dose mapping, for precise reporting of applied doses.85,86 Dose accumulation is an important and potentially indispensable tool for the future reduction of radiotherapy related toxicities through online adaptive radiotherapy. Several groups have reported the development of software tools for deformable dose accumulation, QA procedures to benchmark dose accumulation and estimate related dose-volume uncertainties.87, 88, 89, 90, 91 Initial studies comparing different implementations of dose accumulation have reported high variation of resulting accumulated dose maps depending on patient- and fraction-specific details.92

Recent studies have highlighted the potential of using MRI-LINACs for continuous, sequential assessment of functional and biological information about the tumor tissue to monitor treatment-induced changes, assess treatment response and if necessary, create response-adapted radiation dose prescriptions.52,93, 94, 95 Before interventions based on functional imaging information can be tested in clinical trials, potential surrogate parameters for RT outcome need to be identified in dedicated studies, then sequences must be optimized and tested in phantoms and patients.96 In addition, thorough evaluation of the parameter reproducibility, repeatability, and accuracy with respect to reference values to be measured on MR-LINACs is required by comparison to diagnostic scanners and test-retest studies.93 Several studies were recently published using the 1.5 T MR-LINAC demonstrating the possibility of using this system for response adaptive MRIgRT.94,95,97, 98, 99, 100

Increased MRIgRT utilization will depend on multiple factors, including cost effectiveness, treatment outcomes, and adequate workflow training. Because adaptive MRIgRT tends to be more time intensive and requires additional quality and safety assurances than CBCT-based RT, fewer patients can be treated each day.32,101 Improving adaptation process efficiency will be vital to increasing the number of patients that can be treated each day. Current innovations in this direction include real-time MR auto-segmentation, auto-contouring and MR-only planning.67,68,102,103 New quality control measures should be developed to support MRIgRT adaptive workflows.104 Cost considerations should not only consider the number of patients treated each day but also consider treatment efficacy and patient quality of life. Cost-effective analyses show that MRIgRT can reduce costs related to radiation-induced complications.105 Future clinical trials will be needed to identify which cancer types and patient subsets would benefit most from online adaptive MRIgRT.