First-in-human technique translation of oxygen-enhanced MRI to an MR Linac system in patients with head and neck cancer

Background and purpose: Tumour hypoxia is prognostic in head and neck cancer (HNC), associated with poor loco-regional control, poor survival and treatment resistance. The advent of hybrid MRI – radiotherapy linear accelerator or ‘MR Linac’ systems – could permit imaging for treatment adaptation based on hypoxic status. We sought to develop oxygen-enhanced MRI (OE-MRI) in HNC and translate the technique onto an MR Linac system. Materials and methods: MRI sequences were developed in phantoms and 15 healthy participants. Next, 14 HNC patients (with 21 primary or local nodal tumours) were evaluated. Baseline tissue longitudinal relaxation time (T 1 ) was measured alongside the change in 1/T 1 (termed D R 1 ) between air and oxygen gas breathing phases. We compared results from 1.5 T diagnostic MR and MR Linac systems. Results: Baseline T 1 had excellent repeatability in phantoms, healthy participants and patients on both systems. Cohort nasal concha oxygen-induced D R 1 signiﬁcantly increased (p < 0.0001) in healthy participants demonstrating OE-MRI feasibility. D R 1 repeatability coefﬁcients (RC) were 0.023–0.040 s (cid:1) 1 across both MR systems. The tumour D R 1 RC was 0.013 s (cid:1) 1 and the within-subject coefﬁcient of variation (wCV) was 25% on the diagnostic MR. Tumour D R 1 RC was 0.020 s (cid:1) 1 and wCV was 33% on the MR Linac. D R 1 magnitude and time-course trends were similar on both systems. Conclusion: We demonstrate ﬁrst-in-human translation of volumetric, dynamic OE-MRI onto an MR Linac system, yielding repeatable hypoxia biomarkers. Data were equivalent on the diagnostic MR and MR Linac systems. OE-MRI has potential to guide future clinical trials of biology guided adaptive radiotherapy. (cid:1) 2023 The Authors. Published by Elsevier B.V. Radiotherapy and Oncology xxx (2023) xxx–xxx Thisisan open

Tumours contain hypoxic regions, resulting from an imbalance between oxygen supply and consumption [1]. Hypoxia is prognostic in head and neck cancer (HNC), being associated with both poor loco-regional control and reduced survival [2][3]. In addition, hypoxia impedes radiation treatment by hindering production of free radicals for tumour DNA damage. Consequently, greater radiation dose is required to cause the same tumour damage compared to normoxic microenvironments [4][5]. There is strong evidence that combining hypoxia-modification with radiotherapy improves loco-regional control [6] and overall survival in patients with HNC [7]. Furthermore, hypoxia limits the effectiveness of chemotherapy and immunotherapy [8].
Identifying, mapping and quantifying tumour hypoxia before treatment and following re-oxygenation during radiotherapy may improve stratification of patients based on their hypoxic status. This may assist biology guided adaptive radiotherapy [3,[9][10]. Invasive oxygen electrode measurements, gene signatures [11], immunohistochemistry (IHC) based biomarkers [12] and endogenous blood-based biomarkers [13] can identify hypoxia but do not provide spatial information, track temporal change, or enable analysis of multiple lesions with distinct biology without serial invasive sampling. These factors are all important when considering personalised radiotherapy [5].
PET imaging can non-invasively measure tumour hypoxia and monitor hypoxic changes through treatment. However, PET requires specialist radiochemistry, expensive radiopharmaceuticals and local expertise, which has hindered widespread clinical adoption [14]. Oxygen-enhanced magnetic resonance imaging (OE-MRI) offers a practical and readily translatable technique to assess oxygenation in normal tissues and tumours with spatial resolution superior to PET [15]. In OE-MRI, change in longitudinal relaxation rate (R 1 ) of blood and tissues is measured following inhalation of 100% oxygen or carbogen [16]. Inhaled oxygen molecules dissolve in the blood plasma and interstitial fluid, inducing an increase in R 1 (DR 1 ) via a paramagnetic contrast effect [17]. Studies on preclinical and clinical diagnostic MR systems show that OE-MRI can identify, quantify and map hypoxia in animal and human tumour types and track changes induced by radiotherapy [17][18][19]. R 1 based OE-MRI is distinct from Blood Oxygen Level Dependent (BOLD) imaging which measures changes in the haemoglobin associated effective transverse relaxation rate (R 2 *) [20] and is susceptible to magnetic field inhomogeneities from air-tissue interfaces found in head and neck anatomy.
OE-MRI is an attractive technique to monitor tumour oxygenation on hybrid systems that combine real-time MRI with radiotherapy delivery. These systems facilitate personalised biology guided adaptive radiotherapy by targeting hypoxic tissue through dose painting or other techniques [9]. However, there are several expected challenges in translating such techniques from diagnostic systems to an MR Linac system. These include hardware differences associated with receive coil and gradient systems [21], the requirement for ancillary equipment within the radiotherapy bunker [22][23] and demonstration that OE-MRI biomarkers derived on the MR Linac are comparable to those derived on diagnostic systems. The aims of this work were to demonstrate OE-MRI feasibility and repeatability on an MR Linac system in healthy participants and in HNC patients.

Materials and methods
Scanner hardware OE-MRI was established and implemented on two systems; a 1.5 T MR scanner used in standard diagnostic healthcare (Philips Ingenia MR-RT system, Philips Medical Systems, Philips MR software version 5.7.1), herein referred to as 'diagnostic MR system' and a 1.5 T MR Linac system (Elekta Unity, Philips MR software version 5.7.1).
On the diagnostic MR system, a 16-channel posterior spinal array, a 32-channel large flexible anterior array, positioned on a coil bridge, and two single channel loop coils positioned laterally, were used. On the MR Linac system, a 4-channel posterior array and a 4-channel anterior array, positioned on a coil bridge, were used. Imaging on both systems was performed on a flat table-top to mimic radiotherapy setup.

Imaging protocol and equipment
Imaging sequences were optimised in phantoms and healthy participants on the diagnostic MR system. Finalised sequences were replicated as close as possible on the MR Linac system. A schematic of the imaging protocols ( Supplementary Fig. 1) and sequence parameters (Supplementary Table 1) are provided.
All imaging was acquired in the transverse plane without slice gaps. Healthy participants and patients all initially breathed medical air during acquisition of a T 2 w fast-spin echo (FSE) anatomical image and a baseline 3D inversion recovery turbo field echo (IRTFE) T 1 map with five inversion pre-pulse delays. This was followed by the 3D dynamic IRTFE sequence (OE-MRI acquisition with 12 s temporal resolution, consisting of initial scans 1-25 of the IRTFE dynamic sequence; scan time 5 minutes). Then, gas delivery was switched to 100% O 2 for scans 26-70 of the dynamic sequence (9 minutes), before returning to medical air breathing on scans 71-91 (4 minutes). Following this, a 3D fat-saturated T 1 w FFE was acquired. In HNC patients only, contrast agent (gadoterate meglumine (Dotarem, Guerbet), 0.2 mol/kg at 3 ml/s, 20 ml flush) was delivered to facilitate tumour contouring on the T 1 w FFE sequence. Contrast agent was delivered by contrast power injector (MRExperion, Bayer) installed on both systems.
Modifications were required to the MR Linac system room to install medical air and oxygen from the magnet room supply ports to a low-flow air-oxygen gas blender (Inspiration Healthcare) to deliver either 21% or 100% O 2 . Gases passed through a flowmeter to control flow at continuous 15 l/min and then through flexible tubing to a non-rebreathe oxygen mask (Ecolite TM , Intersurgical Ltd). This circuit ensured similar gas delivery to our diagnostic systems [18], where waveguides allow gas hoses and tubing to be passed between the control room and the magnet room (Supplementary Fig. 2).

Phantom data acquisition
Sequence evaluation was performed using a Eurospin TO5 phantom (Eurospin, Diagnostic Sonar). Prior to imaging, the phantom was left in the diagnostic MR system room for temperature stabilisation at the ambient room temperature of 24°C. The phantom was immediately transferred to the MR Linac following diagnostic MR system measurements to minimise temperature changes in the gel samples between imaging on the different systems. The phantom imaging protocol (Supplementary Fig. 1) consisted of repeated baseline T 1 measurements, carried out 20 minutes apart. In addition, the presence of drift in the dynamic IRTFE measurement was evaluated in the absence of the gas challenge.

Data acquisition in healthy participants and patients
All participants were recruited after research ethics approval and provided written informed consent (ClinicalTrials.gov identifiers: NCT04903236 and NCT03646747).
Initial protocol development was in 4 healthy participants (data not included). The resultant locked protocol was acquired in 6 healthy participants using the diagnostic MR and equivalent protocol was acquired in 5 different healthy participants using the MR Linac system.
The locked protocols were then tested for feasibility in 3 patients on the diagnostic MR system. Then data were acquired in 6 HNC patients using the diagnostic MR system and 5 different HNC patients using the MR Linac system.

Data processing and image analysis
In phantoms, baseline T 1 maps were calculated using a region of interest (ROI) positioned in the central slice of each gel sample. Linear fitting was performed on the dynamic IRTFE data to assess drift. In human subjects, motion correction of dynamic OE-MRI data and registration of pre-gas-challenge (baseline) T 1 mapping and dynamic OE-MRI datasets was carried out for all participants using a deformable registration method in Elastix [24].

Oxygen-Enhanced MRI on an MR Linac
Baseline T 1 maps were derived by non-linear least squares fitting to the signal data (S(TI)) acquired at five inversion pre-pulse delay times (TI). The IRTFE sequence employed a repetition time (TR) > 5 T 1 and very short echo time (TE), such that the TR and TE terms can be ignored. For this reason, the non-linear fit estimation of T 1 reads as: Here, S 0 is the equilibrium signal, TI is the previously described inversion pre-pulse delay time and k is the inversion efficiency parameter. S 0 , T 1 and k were fit as free parameters. Measurement of baseline T 1 then permitted the conversion of dynamic signal change (DSI(t)) to DR 1 (t) (where, DR 1 (t) = R 1,O2 -R 1,air ) [25]. R 1,air was calculated as the median of DR 1 measurements 2-25 acquired during the air phase and R 1,O2 values was calculated as the median of dynamic measurements 60-70 acquired during the 100% oxygen phase.
In healthy participants, DR 1 measurements were obtained from a ROI positioned within the nasal concha and the tongue. In patients, primary tumours and any local metastatic lymph nodes were outlined on T 1 w FFE gadolinium enhanced contrast images by a clinical oncologist (6 years' experience) and transferred to the motion corrected datasets allowing lesion DR 1 estimation. Contouring was carried out using JIM software (JIM 6, Xinapse Systems).
Whole ROI analysis measured DR 1 in the nasal concha and tongue of healthy participants and in tumour lesions in patients. In addition, DR 1 maps were produced from voxel-wise processing to provide spatial representation of the oxygen-induced DR 1 changes within the lesion.

Statistical analysis
Phantom T 1 data was assessed by Bland Altman analysis and limits of agreement (LOA) (cohort mean (l) ± 1.96 standard deviation (r)) are presented [26]. We compared l and r of healthy participants and HNC patient T 1 and DR 1 estimates and derived the repeatability coefficient (RC) and within-subject coefficient of variation (wCV) from repeat measurements [27]. In this small cohort, repeat measurements were deemed to show no significant difference if there was overlap in their corresponding 95% confidence intervals on RC [27][28][29]. Oxygen-induced change in DR 1 was assessed by unpaired t-test between DR 1 measurements on air vs oxygen breathing, p < 0.05 was deemed significant in this work. Data processing and analysis was carried out using MATLAB (Mathworks).

Results
Phantom T 1 measurements were derived (Supplementary Fig. 3) and their repeatability was determined for each MR system and across systems (Supplementary Fig. 4). On the diagnostic MR, the difference in mean T 1 was (l ± r) À4.0 ± 6.2 ms (LOA = -16.2 to 8.1 ms). In comparison, on the MR Linac, the difference in mean T 1 was À6.9 ± 6.6 ms (LOA = -20.0 to 6.1 ms). Agreement between the two MRI systems was assessed (Supplementary Fig. 4) and the difference in mean T 1 was 2.3 ± 22 ms (LOA = -40.0 to 44.5 ms).
Signal drift during the dynamic sequence used for OE-MRI measurements was assessed on both MRI systems. The signal courses for each of four representative T 1 gel tubes were linear, horizontal, and showed no upward or downward trends during the 18 minute acquisition on either system ( Supplementary Fig. 5). Linear fitting to the dynamic IRTFE data in these gels showed the largest DR 1 drift measured were DR 1 = 1.4x10 -5 min À1 s À1 and DR 1 = 6.0x10 -5 min À1 s À1 on the diagnostic MR and MR Linac systems respectively. Collectively, this shows that phantom T 1 measurements are repeatable and free of drift on both the diagnostic MR and the MR Linac.
Next, we evaluated if OE-MRI could be tolerated on the MR Linac and detect oxygen inhalation in normal tissues. Following sequence optimisation and protocol standardisation, in 4 healthy participants, locked down protocols were performed in a further 11 healthy participants (6 on diagnostic MR and 5 on MR Linac). No adverse events were reported. Hyperoxic gas breathing was well tolerated in all participants. Each participant had two scan visits (B1 and B2) evaluating the nasal concha and the tongue, at 13. 4 ± 21.7 days apart. All datasets were deemed suitable for inclusion in analysis, following motion correction and registration.
The cohort DR 1 induced by oxygen challenge was significant in the nasal concha on both MR systems (p < 0.0001) and DR 1 timecourse curves appeared similar in shape and magnitude on both MR systems (Fig. 1, C-D). This indicated that OE-MRI could detect oxygen inhalation on both systems. On the diagnostic MR system DR 1 in the nasal concha was 0.052 (±0.020) s À1 for B1 and 0.062 (±0.029) s À1 for B2 (RC of 0.040 s À1 (95% CI: 0.026-0.089 s À1 ) and wCV of 25%). By comparison, on the MR Linac the DR 1 in the nasal concha was 0.061 (±0.009) s À1 for B1 and 0.055 (±0.013) s À1 for B2 (RC of 0.023 s À1 (95% CI: 0.015-0.057 s À1 ) and wCV of 15%; Fig. 1, E-F).
The cohort DR 1 induced by oxygen challenge was also significant in the tongue measured on the diagnostic MR (p = 0.03) and MR Linac systems (p = 0.0003). However, the magnitude of change was substantially less than that seen in the nasal concha (Supplementary Table 2 and Supplementary Fig. 6).
We then evaluated feasibility in 3 patients on the diagnostic MR system. Clinical details are provided in Supplementary Table 3. In each patient, a statistically significant oxygen DR 1 was detected, resulting in oxygen-induced changes of 0.018, 0.032 and 0.010 s À1 respectively (p < 0.0001) (Fig. 2). This shows that OE-MRI can be performed with dynamic, volumetric IR-based sequences that cover the head and neck region in a clinically practicable timescale.
Next, we evaluated repeatability of OE-MRI in a further 11 patients scanned twice (B1 and B2) using the locked protocol at 5.6 ± 1.6 days apart. Details of patient age, gender, tumour histology, stage and target lesions imaged are listed in Table 1. Again, no adverse events were reported and hyperoxic gas breathing was well tolerated in all patients. Analysis proceeded in 6 diagnostic MR system patients (5 primary tumours; 6 nodal metastases) and 5 MR Linac system patients (5 primary tumours; 2 nodal metastases), resulting in 18 lesions available for evaluation. Sample DR 1 maps are shown in Fig. 3 and show spatial similarity between B1 and B2. All datasets except one (see Table 1; patient 5, tumour) were deemed suitable for inclusion in analysis, following motion correction and registration.
Median values and repeatability of lesion baseline T 1 and oxygen-induced DR 1 are summarised in Table 2 and Fig. 4. OE-MRI DR 1 time-course curves appeared equivalent in shape and magnitude across both MR systems. The cohort DR 1 induced by oxygen challenge was significant for patient lesions on both MR systems (p < 0.0001). This shows that OE-MRI is repeatable and detects tumour hypoxia.

Discussion
The MR Linac presents an opportunity to map and quantify tumour function daily or several times per week, in addition to tracking change in tumour size and position. Quantitative imaging techniques including diffusion weighted imaging (DWI) [30], intravoxel-incoherent motion (IVIM) [31], T 1 and T 2 relaxometry [32] and chemical exchange saturation transfer (CEST) [33] have been described on MR Linac systems. However, few of these techniques yield biomarkers that measure aspects of the tumour microenvironment with proven clinical relevance. In distinction, imaging hypoxia -a well-recognised prognostic and predictive indicator of outcome following radiotherapy -has clear rapid translational potential.
Prior to this study, we and others have demonstrated that OE-MRI can induce signal changes in healthy tissues [16,34], can identify, quantify and map hypoxic sub-regions in mouse [17][18], rat [19], rabbit [35] and human [36][37][38][39] tumours, and can track response to therapy in patients with lung cancer [18]. While our previous work has focused on the value of OE-MRI combined with a perfusion sequence to exclude necrosis [17][18], most of the studies listed above from other groups tend to perform OE-MRI without this additional step. In this work, we have performed a first-in-human translation of OE-MRI as a standalone sequence onto the MR Linac and demonstrated its feasibility and repeatability in HNC patients.
We initially evaluated the OE-MRI protocol using phantoms to confirm T 1 measurement repeatability and assess drift in dynamic measurements. This is important since the baseline T 1 measurement is used to convert the oxygen-induced signal change to DR 1 , which is effectively proportional to oxygen concentration in tissue [20]. We then installed a permanent supply of oxygen and medical air from the main hospital gas supply through the MR Linac bunker to the treatment room. New gas panel interface and gas ports were installed without the need for magnet ramp down.
Following optimisation of MR sequences in a cohort of healthy participants on the diagnostic 1.5 T system, we replicated T 1 mapping and dynamic OE-MRI sequences on the MR Linac. We identified the nasal concha as a valuable reference region to confirm oxygen delivery in healthy subjects. Both diagnostic MR and MR Linac protocols were sensitive to oxygen-induced T 1 change in tissues including the nasal concha. Importantly all aspects of signal evaluation -the shape and magnitude of the dynamic curve, the temporal response to gas challenge, the variation between subjects, the repeatability between baseline scan episodes -were equivalent on the MR Linac and our diagnostic system at 1.5 T, despite differences in receive coil and gradient performancedue to the split gradient system of the MR Linac (slight increase of TR by 0.2 ms and TE by 0.1 ms on the MR Linac was deemed acceptable) [21].
We then translated OE-MRI to evaluate patients with HNC on both the diagnostic MR system, to define system performance, and then on the MR Linac. Again, we showed that the shape and magnitude of the dynamic curve, the within subject variation, the repeatability and the temporal response to gas challenge were equivalent across the two systems. Practical constraints prevented data being acquired from the same participants across both the diagnostic MR and the MR Linac, but despite this data appeared very similar across both platforms.
Data quality was formally assessed. One lesion dataset was corrupted by motion beyond recovery using our registration technique, but otherwise we were successful in applying dynamic, volumetric IR-based OE-MRI in 15 healthy participants (26 scans) and in 14 HNC patients (25 scans) with a success rate of 50/51 (i.e. 98%). Further, DR 1 repeatability was measured for each MRI Oxygen-Enhanced MRI on an MR Linac

Oxygen-Enhanced MRI on an MR Linac
Limitations of our work include the need to shorten the protocol described here to enable a revised version to fit into the MR Linac workflow -from 40 minutes in this work to 15 minutes or less; the need to add in a perfusion sequence such as DCE-MRI [41] to exclude regions of necrosis; potential need for optimisation using thermoplastic shells for immobilisation, as is commonplace for neck radiotherapy, and the need to establish multisite feasibility and reproducibility across distinct geographical sites [42][43].
In summary, we have successfully performed OE-MRI in patients with HNC. We then performed first-in-human application of OE-MRI on a MR Linac. Further work is underway to assess treatment response and dose painting using OE-MRI, and to optimise the technique for real-time biology guided adaptive radiotherapy that has short scan time, rapid analysis and sufficient QA procedures, to be performed while the patient lies in the scanner.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.