Initial Feasibility and Clinical Implementation of Daily MR-guided Adaptive Head and Neck Cancer Radiotherapy on a 1.5T MR-Linac System: Prospective R-IDEAL 2a/2b Systematic Clinical Evaluation of Technical Innovation

Abstract

Introduction:

This prospective study is the first report of daily adaptive radiotherapy (ART) for head & neck cancers (HNC) using a 1.5T MR-linac, with particular focus on safety & feasibility and dosimetric results of an on-line rigid registration-based adapt-to-position (ATP) workflow.

Methods:

Ten HNC patients received daily ART on a 1.5T/7MV MR-linac, six using ATP only and four using ATP with one off-line adapt-to-shape re-plan. Setup variability with custom immobilization masks was assessed by calculating the average systematic error (M), standard deviation of the systematic error (Σ), and standard deviation of the random error (σ) of the isocenter shifts. Quality assurance was performed with a cylindrical diode array using 3%/3mm γ criteria. Adaptive treatment plans were summed for each patient to compare delivered dose with planned dose from the reference plan. The impact of dosimetric variability between adaptive fractions on the summation plan doses was assessed by tracking the number of optimization constraint violations at each individual fraction.

Results:

The random errors (mm) for the x, y, and z isocenter shifts, respectively, were M = −0.3, 0.7, 0.1; Σ = 3.3, 2.6, 1.4; and σ = 1.7, 2.9, 1.0. The median γ pass rate was 99.9% (range: 90.9%−100%). The differences between the reference and summation plan doses were within [−0.61%, 1.78%] for the CTV and [−11.74%, 8.11%] for organs at risk (OARs), though percent increases in OAR dose above 2% only occurred in three cases, each for a single OAR. All cases had at least two fractions with one or more constraint violations. However, in nearly all instances, constraints were still met in the summation plan despite multiple single-fraction violations.

Conclusion:

Daily ART on a 1.5T MR-linac using an on-line ATP workflow is safe and clinically feasible for HNC and results in delivered doses consistent with planned doses.

1. Introduction

High-field magnetic resonance imaging (MRI)-guided radiation therapy (RT) has become a clinical reality with the commercial implementation and health-system accreditation of a hybrid 1.5T/7MV MRI/linear accelerator (MR-linac) [1–3]. The enhanced soft tissue contrast of the MR-linac allows visualization of non-bony target volumes and organs at risk (OARs) during treatment setup, enabling daily on-line treatment plan reoptimization for adaptive radiotherapy (ART) [3–5]. First clinical experiences with the MR-linac have been previously reported for oligometastatic, pelvic, and breast tumors[6–9], but utilization of the device has since expanded to a number of other tumor sites, including head and neck cancer (HNC).

Rather than ad hoc technology development and implementation via unstructured, non-sequential processes, the MR-Linac Consortium [10] has developed the R-IDEAL conceptual framework of systematic technical and clinical reporting processes for clinical radiotherapy applications [11]. Furthermore, to our knowledge, there are no reports of high-field MR-guided daily ART for HNC on a hybrid MR-linac system. Thus, pursuant to this commitment, we sought to undertake an R-IDEAL Stage 2a (“technical optimization of the innovation for treatment delivery”)/2b (“proof of early clinical effectiveness and safety of the innovation”) study with the following aims:

1) Describe an adapt-to-position (ATP) workflow for daily ART using conventional fractionation and ensuring preservation of pre-therapy dose constraints for target volumes and OARs; 2) Demonstrate the capacity of the 1.5T MR-linac system to deliver safe and clinically acceptable ART plans for HNC; 3) Characterize performance based on treatment times, quality assurance (QA) results, setup corrections, cumulative doses, and dosimetric variability between adaptive plans.

2. Methods

2.1. MR-Linac Clinical Workflow for Head and Neck

The clinical workflow described in this paper was developed as a multi-institutional effort through the MR-Linac Consortium’s Head and Neck Tumor Site Group. Although initial implementation at a single site (MD Anderson Cancer Center; Houston, TX) is described, this workflow is based on guidelines iterated through the Tumor Site Group prior to large-scale implementation (Appendix A).

2.1.1. Pre-treatment Workflow

Patient Selection

HNC patients are deemed eligible for treatment on the 1.5T/7MV MR-linac (Unity; Elekta AB; Stockholm, Sweden) if the following criteria are met: Patients must have intact tumor or an MR-visible reference OAR (e.g. skull base) and planning target volume (PTV) length <20 cm in the super-inferior direction (due to field size constraints [12]). Patients with claustrophobia, contraindications to MRI (metal objects, pacemakers, etc.), and/or compromised airway are excluded. Because the MR-linac is located in the out-patient radiation treatment center at our institution, the patient must be ambulatory. If the patient is admitted to the in-patient facilities at any time during treatment, he or she must be treated with a back-up plan on a conventional linac at the in-patient center. The full Patient Evaluation Form and MRI Safety Screening Form are given in Appendices B and C.

Simulation Protocol

All patients receive both CT and MR simulations on the same day and with the same immobilization devices to ensure that the CT and MR images may be rigidly registered [13]. The CT simulation is used for planning for both the MR-linac treatment and the back-up plan on a conventional linac. A tabletop overlay is used on the CT scanner to reproduce the actual couch top in the MR-linac treatment room. Placement of the index bars along the couch establishes the location of patient immobilization aids and the actual tumor volume with respect to the isocenter on the MR-linac. The tabletop overlay is also equipped with a hoop-shaped device representing the maximum clearance of the MR-linac bore to ensure that the patient and all immobilization aids will be able to fit into the MR-linac. The patient is immobilized with a custom thermoplastic immobilization mask and cushion (Klarity Medical Products; Heath, OH) (Figure 1).

Figure 1:

The MR-linac simulation and on-line adaptive re-planning workflow. a) Custom head & neck immobilization mask with the coil on the MR-linac; b) indexing bar on the CT simulation MR-linac table overlay; c) CT sim hoop device used to test clearance in MR-linac bore; d) planning CT with contours; e) reference treatment plan on planning CT; f) daily setup MRI from MR-linac (2-minute T2 protocol); g) rigid registration of planning CT and setup MRI; h) adaptive plan using ATP workflow; i) dose volume histogram (DVH) of reference plan; j) DVH comparison between reference and adaptive plans.

The MR simulation protocol includes two standard vendor-provided non-contrast T2-weighted 3D MRI sequences (2-minute and 6-minute scans; scan protocol parameters provided in Appendix D) and may include additional specialized sequences (diffusion-weighted imaging (DWI), 2D balanced fast field echo (bFFE) in three orthogonal directions for motion estimation, etc.). The physician determines at the time of MR simulation whether the 2-minute T2-weighted scan is adequate for daily setup and plan reoptimization; if the tumor or reference OAR is not clearly visible, the 6-minute scan is used for daily imaging.

Treatment Planning and Pre-Treatment Preparation

The CT and T2-weighted MR simulation images are fused for contouring of the target volume(s) and OARs by the physician and dosimetrist, respectively. Treatment planning is done in Monaco (version 5.4; Elekta AB), which employs a Monte Carlo-based dose calculation engine. All MR-linac plans use step-and-shoot intensity-modulated radiation therapy (IMRT). An initial reference plan is created by a dosimetrist based on the target prescription dose(s) and OAR constraints (Appendix E) provided in the physician’s planning directive. The optimized plan is reviewed by the physician and physicist. Upon approval, the treatment plan is exported to the radiotherapy treatment data management system (Mosaiq; Elekta AB), where a Plan of Care is created for the patient. The Plan of Care specifies the exact MRI exam card(s) to be loaded during each treatment session.

2.1.2. Online Clinical Workflow

The Unity on-line adaptive workflow has been previously described by Winkel et al. [3]. At our institution, a T2-weighted MRI is used for daily setup verification and plan adaptation for HNC patients. The 6-minute scan was used for the first three patients to be conservative, but the 2-minute scan was used for all subsequent patients because the target or reference anatomy could be adequately visualized. The daily setup image is then fused with the reference plan image.

Two on-line adaptive workflows are possible with the Unity system: adapt-to-position (ATP) and adapt-to-shape (ATS). ATP is a virtual isocenter shift of the reference plan based on rigid registration with either a dose recalculation or plan reoptimization. ATS involves deformable image registration to propagate contours onto the daily setup image followed by a full plan reoptimization on the current anatomy [3]. The ATP workflow was used for on-line plan adaptation for six patients, and a combination of on-line ATP and off-line ATS was used for four patients (maximum one ATS during treatment course) to account for soft tissue deformation.

The following acceptance criteria are used for on-line plan review and to determine when an ATS is required: After fusion of the setup and reference plan images, if the isocenter shift is <5 mm in all directions and planning constraints are satisfied, the ATP workflow can proceed. If all shifts are <5 mm but any one dosimetric criterion is not satisfied for three consecutive fractions, the physician must be consulted prior to treatment and an off-line ATS is considered. If any shift is >5 mm, the patient must be repositioned; if any shift is still >5 mm (most likely due to patient weight loss), the physician must approve the plan prior to treatment, and an off-line ATS will be performed.

On-line ATS is challenging for HNC due to the large number of structures used in plan optimization. To improve the efficiency of the on-line workflow, ATS was only performed off-line to create a new reference plan for use in the subsequent days’ on-line ATP. When an ATS is performed, a 6-minute T2-weighted MRI is acquired during the previous fraction as the reference image for the ATS plan because the 2-minute scan is not adequate for target delineation. In the current study, the goal of treatment plan reoptimization was to ensure the delivery of pre-therapy planned doses to target volumes and OARs, referred to as an ART_{ex_aequo}¹ approach in the nomenclature defined by Heukelom and Fuller [14]. Thus, in any off-line ATS’s, the target structures were registered rigidly rather than deformably and were not modified in any capacity; only OAR contours were updated to reflect anatomical changes.

Four methods are available for plan recalculation/reoptimization: “original segments,” “adapt segments,” “optimize weights,” and “optimize weights and shapes” [3]. In our ATP workflow, we use “adapt segments” when the isocenter shift is <1 mm in all directions. Otherwise, we use “optimize weights” first then “optimize weights and shapes” if the plan does not satisfy dosimetric criteria. For ATS, we use “optimize weights and shapes.”

During on-line plan adaptation, motion monitoring with a bFFE sequence is turned on to ensure that the patient has not deviated from the original position. Once the adaptive plan is created, it must be approved by the physicist and physician. Next, a secondary monitor unit (MU) calculation is performed using RadCalc (LAP/LifeLine Software, Inc; Austin, TX) for beamby-beam validation of the MUs in the adaptive plan. Upon final approval, the treatment is delivered. IMRT QA is performed after the treatment using a cylindrical diode array.

2.2. Patient Cohort and Informed Consent

This prospective study included the first ten HNC patients at our institution who satisfied the following criteria: treated with multiple IMRT fractions on the MR-linac, enrolled in the observational clinical trial (NCT04075305; institutional IRB PA18–0341), and not enrolled in any other clinical trials involving non-conventional dosing or fractionations schemes. These patients provided written informed consent for their images, treatment plans, and clinical data to be used.

2.3. Treatment Times, IMRT Quality Assurance, and Setup Variability

Treatment times were calculated from the time stamps in the record & verify system (Mosaiq). Three times are reported: 1) setup and plan reoptimization (the time between the patient’s arrival at the MR-linac and the delivery of the first beam); 2) beam delivery (the time between delivery of the first and last beam); and 3) total treatment time.

IMRT QA was performed for every adaptive plan following treatment. An MR-compatible cylindrical diode array (ArcCHECK MR; Sun Nuclear Corporation; Melbourne, FL) was used with 3% dose/3mm distance to agreement γ criteria [15–17]. We used a tolerance limit of γ≥95% and an action limit of γ≥90%, per consensus recommendations [18]: a γ value of 90% or more was considered passing, but γ values between 90% and 95% were visually inspected prior to passing QA.

Daily setup variability was quantified by recording the isocenter shifts along the x (left/right), y (superior/inferior), and z (anterior/posterior) axes at each adaptive treatment delivered on the MR-linac (253 fractions among 10 patients). Systematic and random errors in each direction were calculated using Van Herk’s method [19], such that M represents the mean systematic error of all isocenter shifts for all patients, Σ represents the standard deviation of the systematic error, and σ represents the standard deviation of the random error.

2.4. Comparison of Reference Plan and Daily Adaptive Plans

In order to validate the daily plan quality generated by the daily ATP workflow, the initial reference plan was compared to the accumulated dose of the daily adaptive plans (summation plan) for each patient. In the Unity ATP workflow, all treatment plans generated from the same reference plan have their dose distributions calculated on the reference image set (i.e. the treatment plans and doses are already in the same frame of reference). This enables rigid dose summation of the ATP plans, whereby all daily adaptive plan doses are scaled to a single fraction then added voxel-by-voxel. In this study, daily adaptive plans were summed using rigid dose summation on the reference image set to create the summation plan. When ATS was performed, the ATS reference image was rigidly registered to the initial reference image prior to plan summation.

Plans were evaluated using the actual values of the quantities used as dosimetric constraints for each patient, such as maximum dose (D_max) to the spinal cord or mean dose (D_mean) to the parotid glands (Appendix E). In cases where there were multiple prescription dose levels to different target volumes, only the highest dose was considered for analysis. The summation plan and reference plan for each patient were compared by calculating the percent difference of each plan quality metric:

$$\%\text{ difference }=\frac{D_{\text{sum }}-D_{\text{ref }}}{D_{\text{ref }}}\times 100\%$$

In cases where the backup plan was delivered on a conventional linac for one or more fractions, only the MR-linac plans were considered for dosimetric analysis by scaling the reference plan to the number of fractions delivered on the MR-linac prior to taking the percent difference. The cumulative dose at each fraction was calculated for one representative case and compared to the expected dose (the cumulative dose if the initial reference plan had been delivered at each fraction). The number of times an adaptive plan violated a dosimetric constraint was also tracked for each patient.

2.5. Accuracy of Deformable Image Registration in Adapt-to-Shape Workflow

To evaluate whether an on-line ATS workflow would be feasible for future patient treatments, we performed a post-hoc analysis of the quality of contours propagated via the “adapt anatomy” deformable image registration tool available through Monaco’s ATS workflow. For each case, select OAR contours (bilateral parotid glands, bilateral submandibular glands, spinal cord, brainstem, and mandible) from the CT simulation were deformed onto the daily setup MRI for fractions 1, 10, 20, and 30. The OAR contours were also rigidly propagated from the CT simulation to the MR simulation, and the contours were deformed from the MR simulation to the same daily setup MRIs to test the quality of MR-to-MR registration. For ground truth comparison, the MRIs were manually segmented by a physician. The deformed contours were compared to the physician-segmented contours using the Dice similarity coefficient (DSC) and mean surface distance (MSD). This analysis was performed on all cases that had at least 30 fractions.

3. Results

3.1. Patient Cohort

The patient cohort included eight males and two females with median age 62 (range: 39–80). The most common treatment sites were larynx (n=3) and oropharynx (n=3). A total of 253 adaptive treatments were delivered on the MR-linac among the ten patients. Seven patients received at least one treatment with their backup plan on a conventional linac due to machine downtime and/or admittance to in-patient facilities. All patients were treated with the ATP on-line workflow, with four patients receiving a mid-treatment off-line ATS re-plan due to anatomical changes. Full treatment information is provided in Table 1.

Table 1:

Patient treatment information for the 10 cases.

Case	Treatment site	Total prescription dose (cGy)	Total no. of fractions	No. of fractions treated on MR-linac	Tumor stage	Concurrent systemic treatment?	Intent of treatment	Adaptation method	Time point & reason for ATS
1	Larynx (true vocal cord)	6300	28	24	T1 N0 M0	No	Curative	ATP
2	Larynx (true vocal cord)	6300	28	24	T1a N0 M0	No	Curative	ATP
3	Oropharynx (base of tongue)	6996	33	26	T4 N2 M0	Yes	Curative	ATP
4	Orbit	1400	4	4	Recurrent	No	Palliative	ATP
5	Hypopharynx (pyriform sinus)	6996	33	25	T1 N0 M0	Yes	Curative	ATP
6	Oropharynx (tonsil)	6996	33	29	T2 N1 M0	Yes	Curative	ATP/ATS	Fx 24*
7	Nasosinus	6000	30	28	T2 N0 M0	Yes	Curative	ATP
8	Larynx (true vocal cord)	6525	29	27	T1b N0 M0	No	Curative	ATP/ATS	Fx 6†
9	Oral cavity (oral tongue)	6996	33	33	Recurrent	Yes	Curative	ATP/ATS	Fx 26*
10	Oropharynx (tonsil)	6996	33	33	T2 N1 M0	No	Curative	ATP/ATS	Fx 8*

Reasons for ATS (last column):

^*Patient experienced shoulder mass loss leading to systematic positioning error/misalignment of rigid registration in the low neck

^†Patient experienced weight loss leading to systematic neck flexion

3.2. Treatment Times, IMRT Quality Assurance, and Setup Variability

The minimum, lower quartile, median, upper quartile, and maximum treatment times (minutes) were 20, 27, 31, 38, 71 for setup and plan reoptimization; 7, 13, 13, 16, 28 for beam delivery; and 31, 41, 46, 53, 85 for total treatment time, respectively. 91% of total treatment times were under 60 minutes.

All adaptive plans in this study passed IMRT QA with a γ value greater than 90% (Appendix F). The minimum, lower quartile, median, upper quartile, and maximum γ values were 90.9%, 99.4%, 99.9%, 100.0%, and 100.0%, respectively. 98% of all adaptive plans had a γ value ≥ 95%, indicating excellent agreement between the expected and delivered dose distributions overall. The outliers below 95% all occurred for a single patient, the first oropharynx patient. The absolute value mean systematic error of the isocenter shifts was ≤ 0.7 mm in all 3 orthogonal directions, and the systematic and random errors were ≤ 3.3 mm (Table 2). The maximum absolute value isocenter shifts in any single fraction were 8.0, 19.7, and 7.3 mm in the x, y, and z directions, respectively. In the y-direction, the shift exceeded 1 cm only four times (Appendix F).

Table 2:

Average systematic error (M), standard deviation of the systematic error (Σ), and standard deviation of the random error (σ) of the x, y, and z isocenter shifts during daily setup of 10 patients (253 adaptive fractions on the MR-linac).

Axis	M (mm)	Σ (mm)	σ (mm)
x (left/right)	−0.3	3.3	1.7
y (sup/inf)	0.7	2.6	2.9
z (ant/post)	0.1	1.4	2.0

3.3. Comparison of Reference Plan and Daily Adaptive Plans

The summed dose to the clinical target volume (CTV) was within [−0.61%, 1.78%] of the reference plan dose in all cases, with the summed dose falling below the reference plan dose in only 2/10 cases (Figure 2). For the contralateral carotid artery, ipsilateral parotid gland, contralateral parotid gland, contralateral cochlea, and brainstem, the summation plan dose was lower than the reference plan dose in all cases. The dose to the ipsilateral cochlea was higher in the summation plan for 3/7 cases, but all cases fell within a maximum dose deviation of 4.3%. Doses to the spinal cord exceeded those in the reference plans for cases 1 and 2 by 8.1% and 2.1%, respectively. However, the total summed doses were between 10 and 11 Gy in these two cases, and these small absolute deviations in dose resulted in higher percent deviations despite the doses falling far below the constraint of 45 Gy. Similarly, dose sparing of 5%–12% was seen for five OARs in case 4, which was the largest dose sparing benefit seen for any case. Case 4 only had four fractions delivered, so the percent differences are larger than measured in most other cases.

Figure 2:

Percent differences between the summation plan and reference plan doses for each plan quality metric specified in Appendix D. Positive values mean that the dose in the summation plan exceeded the dose in the reference plan. The number of treatments delivered on the MR-linac (n) for each case is provided in the legends. In cases where no dosimetric criterion was used for an OAR, that case was omitted from the legend of the graph. “Ipsilateral” and “contralateral” are abbreviated “ipsi” and “contra,” respectively.

All cases had at least two fractions where one or more planning constraint was not met in the adaptive plan (Table 3). However, in most cases where a constraint was violated in at least one fraction, that constraint was still met in the summation plan. For the ipsilateral parotid glands in cases 3, 6, and 9 and the contralateral parotid gland in case 3, the constraints were not met in the initial reference plan and therefore were violated at nearly every fraction and also in the summation plan. There was only one instance (ipsilateral cochlea in case 9) where the constraint was met in the reference plan but not in the summation plan.

Table 3:

Number of times a constraint was violated in an adaptive plan.

	Case 1	Case 2	Case 3	Case 4	Case 5	Case 6	Case 7	Case 8	Case 9	Case 10
Treatment site	Larynx (true vocal cord)	Larynx (true vocal cord)	Oropharynx (base of tongue)	Orbit	Hypopharynx (pyriform sinus)	Oropharynx (tonsil)	Nasosinus	Larynx (true vocal cord)	Oral cavity (oral tongue)	Oropharynx (tonsil)
Total number of MR-linac fractions	24	24	26	4	25	29	28	27	33	33
CTV (high dose)	0	2	2	0	0	2	1	2	3	2
PTV (high dose)	1							6
Parotid (ipsi)	•	•	26*†	0	0	29*†	0	•	31*†	7
Parotid (contra)	•	•	26*†	0	0	0	0	•	2	0
Cochlea (ipsi)	•	•	0	0	0	15	0	•	14†	2
Cochlea (contra)	•	•	0	0	0	0	4	•	0	0
Brainstem	•	•	0	0	0	0	0	•	0	0
Spinal cord	0	0	1	0	0	0	0	0	0	0
Carotid (contra)	0	0	•	•	•	•	•	0	•	•
Brachial plexus (ipsi)	•	•	1	•	•	•	•	•	2	0
Brachial plexus (contra)	•	•	7	•	•	•	•	•	0	0

^•constraint was not used for optimization

^*constraint was not met in initial reference plan

^†constraint was not met in summation plan

Case 9 was chosen as a representative case to illustrate the dosimetric differences between the expected cumulative dose (if the reference plan had been delivered daily) and the actual cumulative dose delivered to the OARs and CTV (Figure 3). Despite the fluctuations in dose between adaptive treatments, minimal deviation was observed between the delivered and expected doses from fraction to fraction (within [−122.7 cGy, 152.4 cGy] over all cumulative fractions). The largest deviation occurred for the ipsilateral cochlea, which had a total of 14 constraint violations out of 33 fractions, resulting in a total dose deviation of 141.8 cGy (4.3%) at the end of treatment. Overall, doses to all OARs were highly consistent between the reference plan and summation plan in all cases, despite constraint violations and variations in daily dose distributions due to on-line plan adaptation.

Figure 3:

The absolute difference between the cumulative delivered dose and the expected delivered dose if the reference plan had been delivered at each fraction for case 9 (treatment site: oral tongue). Positive values mean that the cumulative delivered dose exceeded the reference plan expected dose. An off-line ATS was performed between fractions 25 and 26. This case was chosen as a representative case because all fractions were delivered on the MR-linac.

3.4. Accuracy of Deformable Image Registration in Adapt-to-Shape Workflow

Figure 4 shows the performance of CT-to-MR and MR-to-MR deformable image registration from the CT or MR simulation image to the daily setup MRIs at fractions 1, 10, 20, and 30. The median DSC value fell below the threshold value of 0.8 recommended by the AAPM Task Group 132 [20] for CT-to-MR in all cases. The median DSC met this threshold for MR-to-MR in most cases except for the mandible at fractions 10, 20, and 30. However, the MR-to-MR contour quality was highly variable, with minimum values consistently below 0.6 for all OARs except the brainstem. The median MSD fell within the recommended threshold of 3 mm for all MR-to-MR contours and for the brainstem, submandibular glands, and mandible for CT-to-MR. High variability in MSD was observed for the parotid glands and spinal cord, especially for MR-to-MR. The performance of the deformable image registration did not appear to degrade at later time points for either CT-to-MR or MR-to-MR.

Figure 4:

a) Dice similarity coefficient (DSC) and b) mean surface distance (MSD) between physician-drawn contours and contours propagated via deformable image registration from the CT and MR simulation images to the setup MRIs at fractions 1, 10, 20, and 30.

4. Discussion

ART has begun to play an increasingly prominent role in the treatment of HNC for two primary reasons: 1) significant tumor shrinkage, weight loss, and large anatomical deformations of OARs are often observed for HNC patients during RT [21–23]; and 2) treatment-related side effects from conventional RT can be particularly debilitating for HNC and often persist for years after treatment [24,25]. Several studies have demonstrated clinically significant reductions in doses to the parotid glands [26–28], spinal cord [28,29], and swallowing-related structures [30,31] with one or more off-line plan adaptations over the course of radiotherapy. Because off-line adaptive re-planning is time- and resource-intensive, many studies have focused on identifying anatomical and dosimetric guidelines to determine when plan adaptation would be optimally effective [29,30,32]. However, the clinical introduction of the 1.5T MR-linac has made daily ART for HNC a feasible clinical reality.

In this study, we present our initial experience with daily MR-guided ART for HNC on a 1.5T MR-linac. We employed an ART_{ex_aequo} approach [14] intended to deliver prescription and OAR doses defined prior to therapy. Target volumes were not modified, and OAR contours were updated only when a new reference plan was created via an off-line ATS plan adaptation. We used this conservative approach for our first HNC case series on the MR-linac to test safety and feasibility and to iteratively improve our clinical workflow. However, following the present R-IDEAL stage 2a/2b systematic evaluation, we plan to transition to ART approaches intended to spare OAR doses and/or handle shrinking tumor volumes in the near future [33]. The technological innovations with the MR-linac platform, including enhanced visualization of soft tissue and the on-line ATP and ATS workflows, have the potential to overcome many of the time and resource limitations associated with OAR-sparing and target volume-modifying ART approaches.

Total treatment times ranged from 31–85 minutes, with 91% completed within 60 minutes. Most of the upper outliers were due to isocenter shifts being too large or the on-line adaptive plan not meeting acceptance criteria. In these cases, the patient was repositioned and may have been given a short break out of the mask. Although these treatment durations are longer than HNC treatments on conventional linacs, they are comparable to other specialized image-guided procedures in our clinic.

One common concern with on-line adaptive re-planning is that a plan must be delivered before patient-specific IMRT QA is performed, in significant contrast to conventional QA workflows. However, all adaptive treatment plans in this study passed IMRT QA (γ>90%), with 98% of plans scoring 95% or higher. These high pass rates may be attributed to the comprehensive series of safety checks that are implemented at various stages of the clinical workflow, including annual, monthly, and daily linac and MRI QA as well as the secondary MU check during on-line plan adaptation. This workflow enables any dosimetric issues to be caught long before the patient is treated and ensures that on-line adaptive treatments are safe for all patients. The outliers with pass rates between 90% and 95% all occurred for the same patient, the first oropharyngeal cancer case. Of all treatment sites in this study, the oropharynx cases had the most complex plans. However, all adaptive plans for the next two oropharynx cases had pass rates above 99%, suggesting that plan quality improved with experience.

The setup variability results indicate that immobilization in a custom thermoplastic head and neck mask resulted in highly reproducible patient positioning with setup errors consistent with values previously reported in the literature [34–36]. In both the ATP and ATS workflows of the Unity system, an adaptive plan cannot be created if the isocenter shift exceeds 5 cm in any one direction because the table can only be moved longitudinally along the y (superior/inferior) axis. All isocenter shifts fell far below this threshold, with maximum shifts of 0.80, 1.97, and 0.73 cm in the x, y, and z directions over all 253 adaptive plans. Although shifts greater than 1 cm seem quite large with a mask, this occurred only four times and only in the y direction, which can be attributed to errors in positioning the mask along the index bar. In these cases, a decision was made not to reposition the patient because an acceptable ATP plan could still be generated. However, while on-line plan reoptimization directly accounts for inter-fraction positional variations, minimizing these variations can help the optimization algorithms achieve “better plans” that meet all dosimetric constraints.

One advantage of treating with the MR-linac is the ability to observe motion during the treatment with the motion monitoring cine imaging. Although not directly quantified in this paper, the intra-fraction motion was observed during each treatment and appeared minimal with our immobilization setup and consistent with previous findings [37,38]. The current version of the MR-linac does not allow for intra-fraction motion correction, but the beam can be manually stopped if the target has moved substantially. Large target motion does not appear to be a major issue in the head and neck given that the patients are masked. The only motion observed was due to respiration and swallowing, which, as demonstrated by Bruijnen et al [37], do not typically cause net tumor displacement over time. An example video of an oropharyngeal cancer patient (case 6) midway through treatment is shown in Appendix G.

Differences between the reference plan doses and summed doses were minimal, with the largest percent deviations occurring when the absolute doses were small (as in case 4 when only 4 fractions were delivered or cases 1 and 2 when the spinal cord doses were around 10 Gy). It is important to note that the ATP workflow was used for the majority of fractions, with only a single ATS fraction for cases 6, 8, 9 and 10. There was no modification of the CTV or intentional dosimetric sparing to any OARs for these cases. Any dosimetric sparing to the OARs is likely caused by small changes in the locations of “hot spots” in the dose distribution within each OAR, which is a result of both daily plan adaptation and the inherent randomness of the Monte Carlo dose calculation. Instead of the “hot spot” being located in the same voxel at each fraction, the “hot spot” is smeared around a larger volume over all fractions, effectively reducing the maximum or mean cumulative dose to any OAR compared to the reference plan. Overall, the consistency between the reference and summation plan doses demonstrate that our conservative initial clinical implementation of the ATP workflow on the MR-linac for HNC is safe for patients and performs at least as well as the standard of care from a dosimetric perspective.

One drawback of this dosimetric analysis was that the adaptive plans were summed using rigid plan summation on the anatomy of the reference plan in the ATP workflow. This dose summation approach was necessary because there are currently no validated deformable dose accumulation tools for the Unity system. In the absence of such a tool and with only ten cases in this initial feasibility study, it is difficult to draw conclusions about the dose sparing benefit of this MR-guided ART_{ex_aequo} approach. However, the current standard of care in RT is to plan multi-fraction treatments based on the obviously erroneous assumption that the anatomy remains constant throughout all treatment fractions. Thus, all dose estimates in conventional RT are based on the simulation image, which is analogous to the reference plan anatomy in MR-guided daily adaptive RT. Because all doses were summed on the reference plan anatomy, we may conservatively conclude that the dosimetric variability between daily adaptive plans results in minimal variation from the reference plan over the course of RT.

Another major goal of this study was to characterize the frequency of plan constraint failure for standard head and neck radiation plans and to determine reasonable criteria to accept or reject an on-line adaptive plan based on the pattern of these constraint violations throughout a treatment course. We have shown in this manuscript that most constraint violations are random and do not have a major impact on the cumulative delivered dose. We have also shown that when constraint violations are systematic (i.e. the same constraint is violated in multiple consecutive fractions), our off-line ATS approach allows us to modify the cumulative dose to ensure that the same plan parameters are met without having to keep the patient on the treatment table and in the mask for a time-consuming on-line ATS.

Ultimately, the results presented in this manuscript have enabled our clinic to move toward a protocol with “soft” constraints rather than “hard” constraints, whereby each constraint has a 3% tolerance when creating on-line adaptive plans. Values above 3% overdose for OARs or below 3% underdose for targets are rejected and require modification of IMRT constraints and plan reoptimization. Values that violate the stated constraint but fall within the 3% tolerance are accepted for a single fraction but flagged; multiple consecutive values within this range trigger an off-line ATS. This method allows us to differentiate between random violations that are unlikely to affect the cumulative dose and systematic violations that require off-line plan modification.

In this initial feasibility study, our conservative implementation was to use an on-line ATP/off-line ATS workflow to limit the patient time in mask during each fraction. The greatest bottleneck in the ATS workflow is the time spent manually correcting contours propagated via the “adapt anatomy” deformable image registration tool within Monaco, which is a “one-size-fits-all” tool with no allowable parameter optimization. To test the feasibility of performing on-line ATS in the future, we measured the performance of contours propagated from the simulation images to the daily setup images at fractions 1, 10, 20, and 30. We tested both CT-to-MR and MR-to-MR because we routinely use CT for our reference plans but were interested in seeing whether using MR for the reference plans would improve the deformable image registration for ATS plans. While the median DSC and MSD values were marginally better for MR-to-MR compared to CT-to-MR, the variability in performance across all cases indicates that the image registration is not reliable enough to routinely perform ATS on-line without extensive manual correction. We also tested the image registration quality at four time points to investigate whether the performance decreased as anatomical changes became more significant over time, and if so, whether we could determine a maximum time interval between ATS fractions. No clear trends were observed with time, and DSC and MSD consistently failed to meet the recommended tolerances even at the first fraction. Furthermore, due to the large number of OAR contours and the poor performance of the DIR, our dosimetrists report spending 1.5–2 hours on average for each off-line ATS plan, with the majority of that time spent on editing OAR contours after the image registration. Based on these results, we will continue to implement ATS off-line until a more robust deformable image registration tool can be implemented into the on-line clinical workflow.

Despite these limitations, this effort represents the first report of 1.5T MR-guided daily ART for HNC and the first prospective series of dosimetric analyses of HNC treatment on a hybrid MR-linac system. Consistent with R-IDEAL, in this planned “prospective small uninterrupted case series,” we have demonstrated “technical improvements, feasibility, and safety” [11] of 1.5T daily ART_{ex_aequo} treatment for HNC. The continued R-IDEAL programmatic integration of this data with extant efforts across the MR-Linac Consortium is designed to ensure that rather than ad hoc reportage, iterative technical developments on a new technology are reported systematically. Future prospective analyses combining HNC data from other MR-Linac Consortium sites using the same workflow are currently being planned. Furthermore, first technical studies have shown that MR imaging using the 1.5 T Unity system not only provides excellent anatomical image data [39,40] but also allows for robust and reproducible quantitative imaging techniques, such as DWI [40]. Consequently, functional MR data acquired sequentially during MR-guided RT needs to be investigated systematically in order to pave the way towards future biologically individualized RT of HNC. Next steps include an R-IDEAL Stage 2a evaluation of functional imaging integration and feasibility with DWI, and Stage 2a implementation using an ART_reduco [14] approach within the prospective Phase II randomized MR-ADAPTOR trial [33].

5. Conclusion

The aim of this article was to describe our institution’s clinical workflow for treating HNC on a 1.5T MR-linac and to demonstrate safety and feasibility through a prospective analysis of ten cases treated with an on-line ATP workflow. Treatment times remained under one hour in 91% of cases. All treatment plans passed IMRT QA, and patient immobilization in a custom head and neck mask resulted in highly reproducible patient setups. The daily adaptive plans were very consistent with the reference plans with minimal dosimetric differences. In summary, daily MR-guided adaptive RT for HNC is safe and clinically feasible with the MR-linac, although further studies with more advanced dose accumulation strategies are required to investigate the true dosimetric impact of this treatment approach.