Realizing the potential of magnetic resonance image guided radiotherapy in gynaecological and rectal cancer.

CT-based radiotherapy workflow is limited by poor soft tissue definition in the pelvis and reliance on rigid registration methods. Current image-guided radiotherapy and adaptive radiotherapy models therefore have limited ability to improve clinical outcomes. The advent of MRI-guided radiotherapy solutions provides the opportunity to overcome these limitations with the potential to deliver online real-time MRI-based plan adaptation on a daily basis, a true “plan of the day.” This review describes the application of MRI guided radiotherapy in two pelvic tumour sites likely to benefit from this approach.

The authors Susan Lalondrelle and Shreerang Bhide contributed equally to the work.
intRoduction Multiple challenges exist in radiotherapy (RT) delivery for gynaecological and rectal targets. The target consists of volumes encompassing the primary tumour and elective nodal regions, which are difficult to visualize on CT and move independently of each other. Tumour targets are highly mobile deformable structures and are influenced by adjacent rectal and bladder filling, which is difficult to standardize throughout treatment. Substantial tumour regression can occur, which results in normal tissue falling into high dose regions, and extended field treatments are susceptible to rotational setup error. 1 Intensity modulated radiotherapy (IMRT) reduces dose to normal tissue in gynaecological and rectal RT, 2,3 but tight conformity and sharp dose gradients mean that adequate planning target volume (PTV) safety margins to account for geometric uncertainty are essential to avoid a geographical miss.
The current PTV margins applied to targets are based on margin recipes that aim to ensure 95% of the prescribed dose is delivered to 99% of the target volume, 4 or 95% of the prescribed dose is delivered to 100% of the target volume in 90% of patients. 5 Significant interpatient variability in target motion results in population-based margins that are much larger than necessary in most patients and still miss the target in a small number of cases. The alternative to large margins and increased normal tissue dose is to individualize margins and implement adaptive treatment strategies.
RT is currently planned on a single CT data set obtained at treatment simulation. This may not reflect target and organ at risk (OAR) geometry at the time of treatment delivery. Adaptive radiotherapy (ART) uses information from imaging acquired before or during treatment delivery to modify the treatment plan based on changes in individual target and OAR geometry and biology. Adaptive strategies are classified based on their timescale relative to patient treatment. 6 Offline strategies occur between treatment fractions and typically involve a single or multiple replans. Online adaptation is based on imaging acquired immediately prior to treatment and can be used daily or intermittently. In online adaptation, tumour target and OAR interfraction changes are accounted for, which means that PTV margins can be significantly reduced. 7 Adaptive strategies can also use information from previous treatment imaging to track the actual dose delivered to the tumour target and OARs and correct for any discrepancy between the planned and delivered dose distributions. 8

RAtionAle FoR MRi-guided AdAPtive RAdiotheRAPy (MRigARt) in gynAecologicAl And RectAl cAnceR
MRI is the imaging modality of choice for diagnosis and staging in gynaecological and rectal cancer where it characterizes tumour and local macroscopic extent to inform treatment decisions, assess treatment response and detect recurrent disease. 9-11 It is essential in identifying patients for radiation treatment, determining the radiation treatment field extent and accurate definition of the tumour target from bladder, sigmoid and small bowel.

MRI improves target localization
Target volume delineation on the planning CT in both gynaecological and rectal tumours is difficult because it is not possible to discriminate between tumour and normal tissue. Figures 1  and 2 illustrate improved soft tissue contrast seen on MRI compared to CT for RT treatment planning in rectal and cervix cancer. Compared to CT, target volume delineation on MRI results in significantly smaller rectal and cervix volumes 12,13 and low interobserver variability 14,15 .Studies evaluating interand intraobserver variability in contour delineation on MRI in gynaecological and rectal RT are illustrated in Tables 1 and  2. [12][13][14][15][16][17][18][19][20][21][22][23] In rectal RT, MRI delineation results in significantly reduced tumour length, width and distance of the proximal tumour edge to the anal verge p < 0.05. 12 When gross tumour volume (GTV) is subdivided into tumour located in the sigmoid, rectal and anal sub regions, coverage of the CT contoured GTV was inadequate for tumours with MRI evidence of sigmoid or anal invasion. 21 In cervix cancer, geometric studies show that agreement between target volumes delineated on transverse and para-transverse planes of MRI is good with conformity index 0.71-0.72. 19 In dosimetric studies, overestimation of tumour width on CT results in significant differences in the volume treated to the prescription dose or higher. 13,24 Compared to the CT-based imaging RT workflow, MRIgRT will provide superior visualization of the target and normal tissue immediately before and during treatment delivery. Table 1 and 2 summarizes the published data for contour delineation on MRI in cervix and rectal cancer.

MRI for motion assessment
Extensive target motion occurs in gynaecological and rectal RT and has been reviewed previously. 25,26 With RT for cervix cancer, the primary clinical target volume (CTV) includes any visible tumour, cervix, uterus, upper vagina and parametrium. The elective nodal CTV includes the pelvic and common iliac lymph nodes (LN) and the para-aortic LN in high-risk disease. Motion is largest at the uterine fundus and studies report maximum interfraction motion of over 30 mm. 27 In one study, margins of 15 mm to the primary and nodal CTV failed in 32% of patients and margins of up to 30 mm were required to ensure coverage in 95% of fractions. 27 With RT for rectal cancer the primary target volume includes the tumour and mesorectum, and the elective nodal volume includes the pelvic LN. The entire circumference of the rectum at the level of the tumour is included, because it is not possible to distinguish tumour from normal rectal tissue on CT. The anterior and lateral rectal wall move more than the posterior wall and motion is larger in the middle and upper rectum compared with the lower rectum. 28 Maximum motion occurs anteriorly, particularly in the upper mesorectum, and anterior PTV margins of 24 mm in the upper mesorectum and 15 mm in the lower mesorectum have been recommended. 29,30 Tables 3 and 4 summarize the published data for cervix and rectal interfraction target motion. 27,28,[31][32][33][34][35][36][37][38][39][40][41][42][43] Bladder and rectal filling influence target motion in gynaecological and rectal RT. With cervix treatment, bladder volume is correlated with superior/inferior uterine motion and rectal volume is correlated with cervix and vaginal anterior/posterior motion. 33 With rectal RT, deformation of the mesorectum is largely driven by changes in rectal volume. 29 In both cervix and rectal RT, there is significant interpatient variation in bladder volume despite bladder filling protocols, and both bladder and rectal volumes reduce during treatment. 27,28,34,44 Laxatives may not significantly reduce target anterior/posterior motion from rectal volume variation, because passage of gas can still cause significant target displacement. 37 Figure 3 illustrates CTV positional changes related to bladder volume as seen on CBCT during cervix RT. MRIgART will facilitate implementation of margin reduction through adaptive strategies that account for these geometric changes.

MRI for anatomical response assessment and dose escalation
Significant tumour regression is observed during cervix and rectal RT. 31,34,45,46 In 20 cervix patients having weekly MRI during chemoradiotherapy (CRT), average tumour volume reductions of 59.6% at week 4 were observed, which resulted in increased uterine motion, substantial changes in tumour position and movement of normal tissue, particularly small bowel, into the high dose region. 47 Repeat MRI and planning after delivery of 30 Gy found that a second IMRT plan significantly reduced the volume of bowel irradiated if the primary GTV decreased >30 cc. 47 In a study of 15 rectal cancer patients, mean tumour regression of 46.3% was seen on MRI by week 5 of CRT and regression was fastest in the first 3 weeks of treatment. 45 A further study in 13 patients found that the majority of patients who had a good response to treatment had volume reduction and fibrotic changes during weeks 1-3. 46 There is a move towards organ preservation in rectal patients with a complete radiological response to spare morbidity from surgery. 48 Patients who respond to CRT are more likely to benefit from dose escalation to increase the rate of pathological complete response (pCR) 46 and early assessment to identify these patients is therefore important. Response to neo-adjuvant CRT is dose-dependent with dose escalation of >60 Gy resulting in increased rates of pCR and acceptable toxicity. 49 Tumour boost volume delineation on the initial RT planning CT does not take account of tumour regression during treatment. Repeat imaging during treatment could help select patients who would benefit from radiation dose escalation and would produce more accurate and smaller boost volumes, facilitating increased tumour dose without increased OAR dose and toxicity. 50 Greatest motion of rectum in upper

MRI for biological response prediction and dose delivery assessment
Functional MRI with diffusion-weighted imaging (DWI) and dynamic contrast enhancement (DCE) may predict biological response in rectal and cervix RT and identify patients for dose escalation. 14,51 MRI has potential to act as a biomarker, identifying good and poorly responding tumours to select patients for dose adaptation in order to improve treatment outcomes. 52-55 Studies suggest that DWI can predict pathological complete response early in rectal RT, 53,54,56,57 but there are limitations to the current evidence preventing its routine implementation in patient selection for dose escalation. Most studies were small and did not prospectively determine MRI criteria to differentiate between complete and non-complete response to treatment. Retrospective identification of these parameters introduces selection bias. There was variability in the time points at which imaging was acquired and surgery was performed. For example, patients classified as achieving a non-pCR at 6 weeks following CRT may have been classified as a pCR if surgery was performed at a later date and meta-analysis reports 6% increase rate of pCR with an interval of greater than 6 weeks from the end of preoperative CRT. 58 In cervix RT, DCE and DWI MRI may predict response to CRT and identify patients for dose escalation. 51 Increasing apparent diffusion coefficient (ADC) values from DWI acquired during treatment can detect early signs of treatment response. 55 DCE MRI during treatment detects tumour perfusion. 59 Persistently low perfusion during CRT is correlated with treatment failure and patients with increases in perfusion during CRT have better outcomes. 59 This could identify patients for dose escalation to hypoxic regions, which should increase tumour shrinkage prior to brachytherapy, which we know improves local control. 60 There was however, no technical standardization in these studies, which limits assessment of reproducibility and generalizability. The optimal time to assess biological response and adapt treatment based on these finding has yet to be determined.
MRIgRT will also provide quantitative knowledge of the actual delivered dose and the impact of radiation dose on tumour and normal tissue. This would enable dose compensation strategies and tumour and normal tissue radiobiological modelling.
AdAPtive RAdiotheRAPy (ARt) StRAtegieS 1. Target volume modification based on individual internal motion PTV modification based on data from setup and internal target motion acquired from planning or previous treatment, allows safe reduction of generic population based margins. This is also referred to as a composite volume technique. The range of target motion is modelled during the planning stage or first treatments to generate an internal target volume (ITV). The treatment plan is optimized offline and applied to subsequent treatments. Individualized ITVs in cervix RT account for the range of cervix and   uterine motion with variable bladder volume and may be based on variable bladder filling CT scans acquired at simulation, or using bladder geometry as a predictive tool. 61,62 Compared to population-based margins, individualized margins reduce CTV-PTV margins by 48% (±6%), and bladder and rectal volume within the PTV is reduced by 5-45% and 26-74% respectively. 62 For rectal cancer, an average CTV can be acquired from the radiotherapy planning (RTP) CT and repeat CTs during the first week of treatment. 30 Adaptation after day 4 resulted in a 7 mm reduction in the maximum required PTV margin from 24 to 17 mm and a significant reduction in PTV and dose to the small bowel. 30

Online plan selection strategy
Online plan selection uses imaging acquired at treatment to select a plan from a library of treatment plans generated from multiple PTVs. In cervix RT, evaluated strategies include a plan library using individualized PTVs based on CTV position at different bladder volumes, or PTVs created by the application of incremental margins to the CTV as seen on RTP CT acquired with a full bladder. 62,63 Compared to a standard population margin approach, plan selection results in significantly better target coverage and OAR sparing. [62][63][64] Adaptation based on variable bladder filling CTVs enables reductions in PTV margins from 38 to 7 mm and better CTV D98% > 95% in comparison to the non-ART approach where 17% of treatment fractions have inadequate target coverage. 62,64 When using an incremental margin approach, a 5 mm margin of the day plan could be used in 25% of fractions. 63 Libraries based on variable bladder filling do not account for rectal filling variation or the passage of gas, which are difficult to predict and can significantly influence cervix motion. 65 In rectal cancer, target motion is influenced more by rectal than bladder filling, so a library of plans strategy based on variable bladder volumes is not appropriate. Instead plan selection has been based on plans with variable PTV margins between −25 and +25 mm applied to the anterior CTV, which is where largest variation is seen. 66 This reduced dose to the bladder and small bowel OARs, although the absolute reductions were small. 67 Plan selection in rectal RT is feasible with good plan selection consistency between observers of 75%. 66 Plan selection in both cervix and rectal radiotherapy is being implemented clinically, but is limited by the image quality of CBCT. MRIgRT would facilitate target and OAR localization for online plan selection.

Plan reoptimization
The optimal strategy to account for target and OAR motion and deformation, anatomical and biological response, is to generate a new plan with full reoptimization. This determines the dose distribution based on target and OAR geometry and/or physiology at the time of treatment delivery. 6 A number of planning studies in cervix RT have simulated the benefit of online replanning. 7,68,69 One study of 33 patients compared a 3 mm PTV margin plan without replanning, with an automated weekly replan on real-time patient geometry as seen on MRI. 68,69 Pre-treatment optimization criteria were automatically reapplied to replans without any physics planner intervention. Without replanning, there was a significant reduction in  accumulated dose to the primary CTV, with nine patients failing D98% > 95%. 68 In patients who were replanned, there was a reduction in CTV between 8 and 68% (median 39%) and the D98 CTV constraint was met in all patients. 68 There was no difference in dose to OARs, which might move with the target and remain in the high dose region. This may lead to increased OAR dose in patients where OAR movement is related to the target compared to patients where the OARs move independently. 68,69 A study in 14 cervix patients used 15 mm PTV margins and replanning based on target and OAR geometry on MRI after 30 Gy. 47 There was a reduction in OAR dose with replanning, but in this study the replans were interactively optimized to reflect new anatomy. 47 A planning study to simulate the benefit of online MRIgRT replanned using weekly MRI in 11 patients receiving IMRT for cervix cancer with 4 mm PTV margins. 7 This was compared to plans based on the pre-treatment MRI with primary and nodal PTV margins of 15 and 10 mm. 7 There was a significant reduction in the dose to the bladder, rectum, sigmoid, and small bowel with online replanning. 7

Dose compensation
Adaptation using dose tracking allows reduction in PTV margins because variations in the dose delivered to the CTV compared to the planned dose, can be compensated for in subsequent fractions. The pre-treatment imaging, together with any setup correction applied, is used to determine target and OAR position and the dose delivered at each treatment fraction. This is non-rigidly registered to the planning CT to model anatomical motion and deformation and allows calculation of the accumulated delivered dose. The treatment plan can then be reoptimized to compensate for any problems with dose coverage or to account for adaptation of treatment goals.
Lim et al looked at pre-treatment and weekly MRI in 30 cervix IMRT patients using a 3 mm PTV margin and dose accumulation. 70 They modelled an anatomical driven approach with a single offline replan mid-treatment to account for tumour regression, and a dosimetrically triggered approach if the estimated accumulated D98 to the GTV or primary CTV was low. Without replanning, there was insufficient target coverage in 27% of patients. The anatomical approach improved target coverage and reduced OAR dose, but there were still three patients with insufficient target coverage. Dosimetrically triggered replanning resulted in target coverage in all patients, but no difference in the accumulated OAR dose. 70 Deformable registration is not consistently accurate and validation is difficult. In deformable registration for dose accumulation, particular caution must be taken when tumours have undergone mass change and in areas with sharp dose gradients.

integRAtion oF MRi into RAdiotheRAPy And itS chAllengeS
MRI can be integrated into RT workflow in a variety of ways. In a CT-MRI simulation workflow, the MRI is used for contour delineation at radiotherapy treatment planning (RTP) and the CT provides a robust geometric representation of the patient, an electron density map required for dose calculation and a reference image for patient set up during standard treatment. Any error in image registration will however lead to a systematic geometric error throughout patient treatment. 71 MRI-only simulation reduces potential for image registration error at RTP, but the challenges of geometric distortion and lack of electron density information and material properties inherent to MRI need to be addressed. MRI for RT treatment localization, planning and verification have different demands to those acquired for diagnosis and staging. Specific solutions are required. The main differences relate to patient positioning, image acquisition and sequence parameters and the need for geometric accuracy (Table 5).
A number of MRIgRT technologies are in active development, integrating MRI with external beam RT delivery, providing MRI data immediately before and after treatment, and simultaneously with treatment delivery. [72][73][74][75] They differ in their imaging and treatment adaptation capabilities and their approach to tackling the technical challenges of magnetic and radiofrequency interference and treatment beam transmission through the magnet. Table 6 summarizes the different systems, each presenting advantages and disadvantages. [72][73][74][75][76][77] The MRIdian system (ViewRay Inc, Oakwood Village, OH) has treated over 300 patients since 2014 and integrates a 0.35 Tesla (T) magnet with either three multileaf collimator (MLC)equipped Cobolt-60 heads, or a 6 MV linac. 73,78 The Elekta Unity MR-linac solution (Elekta AB, Stockholm, Sweden) started treating patients in 2017 under pre-CE mark clinical trial protocol. It integrates a 7 MV linac with a high field 1.5 T MRI system from Philips, which uses technology similar to the Philips Ingenia diagnostic systems. 72 Lower magnetic field solutions benefit from a reduction in image artefacts and patient related geometric distortion, and lower energy deposition by the radiofrequency pulses. Higher field solutions benefit from enhanced signal to noise, which improves spatial and temporal resolution and functional imaging capabilities.

technicAl chAllengeS in the ReAliZAtion oF ReAl-tiMe MRigARt
Generation of a new treatment plan based on target and OAR geometry or biology at the time of treatment delivery is the ultimate goal of MRIgART. The main challenge is achieving this in a short amount of time with the patient on the treatment couch. Its clinical implementation is limited by; (1) Requirement for robust automated real-time registration of the newly acquired MRI with the images used for treatment planning.
(2) Requirement for electron density data necessary for dose calculation. In the first clinical applications of MRIgART using the Elekta Unity MR-Linac (Elekta AB, Stockholm, Sweden) and the MRIdian system (ViewRay, Oakwood Village, OH), MRI are acquired immediately before treatment and registered to the reference planning MRI and planning CT using deformable registration. 79,80 Electron density information from the reference planning CT is then transferred to the MRI of the day using the deformation map. 79,80 The standard treatment-planning process requires segmented contours and generates the desired dose distribution from scratch. This is achieved through iterative optimization, driven by defined objective functions set by the planner, which specify the dosevolume constraints for tumour targets and OARs. The planner then fine-tunes the objective functions and repeats the optimization process to further improve the treatment plan by trial and error. This takes too long to be feasibly implemented in real-time MRIgART and faster automated replanning strategies are required.
Segmentation of target and OARs on the daily image is a major challenge in online replanning. Manual segmentation is time consuming and susceptible to inter-and intraobserver variability. Mean time required to manually delineate the pelvic nodal CTV alone is over 30 min, and automated strategies are necessary to reduce segmentation time and improve structure definition. 81 Autosegmentation without prior knowledge uses imaging properties such as voxel intensities and gradients. 82 Alternative strategies incorporate prior knowledge into the segmentation process to improve accuracy and reproducibility and include atlas-based segmentation, statistical shape models, machine learning and hybrid strategies. 82 In atlas-based autosegmentation, an atlas of manually contoured structures is used to propagate structures onto a new data set using deformable registration voxels transformations. [83][84][85] Use of multiple atlases further improves accuracy. 86 Cervix target segmentation on MRI using machine learning results in mean sensitivity and specificity of 85-93% 87 and is faster than atlas based strategies. 88 Accuracy of autosegmentation is not perfect and visual verification is still required. In MRIgART using both the Elekta Unity MR-Linac and the ViewRay MRIdian systems, target and OAR contours are transferred to the online MRI from Daily plan reoptimization does not need to start from scratch and many components of the new plan can be extrapolated from the original fully optimized plan. Plan modification with aperture morphing reduces the number of steps in reoptimization. 90 Segment aperture morphing adjusts the beam segment shape of the MLC, based on the new target position and shape, as seen in the projection from the beam's eye view of each treatment beam. Segment weight optimization can then be applied to improve dosimetry. 90 More complex aperture morphing methods rely on deformable registration. 91,92 Plan adaptation based on previous knowledge from the original plan can also speed up the process. Gradient maintenance strategies maintain the same dose gradient around the target, towards the OARs, as in the original treatment plan. 93 This requires segmentation of the new target but not segmentation of OARS. It may not be suitable for the larger target volumes seen in gynaecological and rectal RT. Interactive dose shaping is based on contoured structures and enables direct manipulation of the initial plan isodose surface shape or the dose to individual voxels. 94,95 Advances in computer power, both graphical processing units and modified central core processing units, can now reduce the time of plan optimization and dose calculation from minutes to seconds. 96,97 Commercial treatment planning systems incorporating advances in adaptive planning are now becoming available.
Plan approval and quality assurance (QA) in real-time MRIgART is challenging. Automation of image acquisition and registration, target and OAR segmentation, treatment dose calculation and adaptive planning optimization is essential in implementing online MRIgART, but creates additional problems. The detailed plan reviews and QA process that occur at pre-treatment during standard RT are not appropriate. Limiting physician plan approval to when plan quality is less than the original treatment plan would improve efficiency. Conventional patient specific QA approaches insert physical phantoms in the treatment beam, which cannot be used with the patient on the treatment couch. An alternative solution is to send the treatment plan to an independent dose calculation engine to verify that the dose distributions agree. 98 Delivery of MRIgRT with the ViewRay MRIdian Cobalt 60 was feasible in 11 rectal patients receiving neoadjuvant chemoradiation with IMRT and simultaneous integrated boost. 99 Daily MRI were acquired for patient setup and verification, and all patients completed treatment. The ViewRay MRIdian has also been used for imaging and RT planning in brachytherapy for cervical cancer. 100 No studies have yet been published for MRIgRT delivery in cervix external beam RT.

concluSionS
MRIgRT in rectal and gynaecological RT will improve all aspects of the treatment workflow. Its most exciting application in gynaecological and rectal RT will be to refine GTV to CTV definition, increased accuracy and precision of target localization for treatment verification and implementation of adaptive strategies to personalize the therapeutic approach. This will facilitate reduced PTV margins and normal tissue irradiation whilst maintaining target coverage. Together with dose adaptation, this will translate into improved tumour control and reduced toxicity for patients. Optimal adaptive strategies need to be determined and challenges remain for the implementation of MRIgART clinical workflow. But technology is exponentially increasing and the ability to personalize and intensify treatment with MRIgART at these tumour sites is no longer an improbable blue-sky ideology but is now within reach.

AcknowledgMent
This work was undertaken in The Royal Marsden NHS Foundation Trust who received a proportion of its funding from the NHS Executive; the views expressed in this publication are those of the authors and not necessarily those of the NHS Executive.