Towards fast online intrafraction replanning for free-breathing stereotactic body radiation therapy with the MR-linac

MRLTP was initially developed in our clinic as an IMRT planning platform able to include the magnetic field into the fluence plan optimization (Bol et al 2012). We have recently upgraded MRLTP to a full planning system including the necessary modules to perform a two-phase IMRT optimization for several linear accelerator (linac) and multi-leaf collimator (MLC) combinations including the MR-linac. It utilizes the research version of the GPU-based Monte Carlo dose engine (GPUMCD, Elekta AB) (Hissoiny et al 2011a, 2011b) which supports calculations under the presence of the magnetic field, along with our custom implementation of the inverse dose optimization presented in Ziegenhein et al (2013), based on minimum and maximum dose prescription per voxel/structure.

MRLTP can be used in a variety of different treatment scenarios by implementing several planning pipelines whose core is ASEQ, a sequencing methodology that iteratively converges to an ideal dose distribution (Kontaxis et al 2015b). Its pipeline enables the inclusion of anatomical changes described by 3D DVFs in a segment-by-segment basis during the plan calculation, by performing a fluence optimization targeting the latest anatomical state in each iteration (Kontaxis et al 2015a). In this work we utilize MRLTP and ASEQ to simulate an inter-beam replanning application targeting the latest anatomical state of the patient, updated in these time intervals.

The experiments were performed on a system with a dual Intel^® Xeon^® E5-2670 v3, 64 GB RAM and two NVIDIA^® GTX Titan X cards for the GPUMCD dose engine, similar in specs with the current industry standard for a treatment planning computer system.

In this section, the different MRI acquisitions and their role in the simulated treatment are presented (figure 1, Imaging).

2.2.1. MRI.

2.2.2. Reference volume and motion estimation.

Treatment planning was performed using the latest specifications of the MR-linac system installed in UMC Utrecht. The installed MLC is fixed at 90 degrees rotation, has 80 leaf pairs with 7.15 mm leaf width and a maximum field size of 57.2 and 22 cm in the perpendicular and parallel to leaf travel direction respectively. A beamlet size of $7.15 \times 2.5$ ${\rm mm}^{2}$ and a dose grid of $3 \times 3 \times 3$ ${\rm mm}^{3}$ were used while all the dose calculations were performed under the presence of 1.5 T transverse field using 3% MonteCarlo statistical uncertainty. The latest 7 MV flattening filter free (FFF) beam model calibrated in our department as described in Wolthaus et al (2015) and a dose rate of 740 MU ${\rm min}^{-1}$ were used. The planning volumes were generated by assigning the density of the body to $1.0$ g ${\rm cm}^{-3}$ (Stam et al 2013).

A single fraction SBRT treatment was simulated where the gross tumor volume (GTV) was prescribed to 25 Gy (Siva et al 2016). For each subject six IMRT beam angles were used depending on the location of the tumor (table 1). The beam angles were selected from the experimental 15-beam configuration in Stam et al (2013), after establishing that they were adequate to provide the required coverage for each subject. Table 2 includes the clinical constraints used, from which the organ at risk (OAR) constraints were proposed in table 3 of Siva et al (2016). For all subjects the same operator performed the planning for both the conventional and adaptive treatments. The GTV volume for the two patients and the healthy volunteer was 1 ${\rm cm}^{3}$ (female, 43 years, solid tumor in the upper pole of the left kidney), 8.5 ${\rm cm}^{3}$ (male, 63 years, cystic tumor in the upper pole of the right kidney) and 4.4 ${\rm cm}^{3}$ (male, 31 years, artificial lesion interpolar of the left kidney) respectively.

Table 1. Beam angles used for the individual cases.

Case	Beam angles
Patient 1	${72^\circ}$, ${96^\circ}$, ${120^\circ}$, ${144^\circ}$, ${168^\circ}$, ${192^\circ}$
Patient 2	${192^\circ}$, ${216^\circ}$, ${240^\circ}$, ${264^\circ}$, ${288^\circ}$, ${312^\circ}$
Volunteer	${48^\circ}$, ${72^\circ}$, ${96^\circ}$, ${120^\circ}$, ${144^\circ}$, ${168^\circ}$

Table 2. Planning constraints.

VOI	Constraints
GTV	99% vol	$>=$	25 Gy
	Max	$<=$	37.5 Gy
PTV	99% vol	$>=$	23.75 Gy
	99% vol	Minimum	22.5 Gy
	1% vol	$<=$	27.5 Gy
Spinal cord	1 ${\rm cm}^{3}$	$<=$	8 Gy
	0.03 ${\rm cm}^{3}$	$<=$	12 Gy
Small bowel	20 ${\rm cm}^{3}$	$<=$	14 Gy
	Full circumference	$<=$	12.5 Gy
Stomach	10 ${\rm cm}^{3}$	$<=$	11 Gy
	5 ${\rm cm}^{3}$	$<=$	22.5 Gy
Skin	Max	$<=$	24 Gy
Heart	15 ${\rm cm}^{3}$	$<=$	16 Gy
Large bowel	As low as reasonably achievable (ALARA)
Ipsilateral kidney	ALARA
Contralateral kidney	ALARA

For all simulations, the calculated plans were delivered to a linac emulator (computer software imitating the physical linac and its functions) (Elekta AB, Stockholm, Sweden) of the currently installed MR-linac in our department. For every delivered beam a log file including the complete machine state (leaf positions, dose rate, delivered MUs etc) every 40 ms, is generated. The log files are then used to match the machine parameters to the timepoints of the dynamic 3D volumes described in section 2.2.2 leading to several segment-to-dynamic volume combinations (Glitzner et al 2015). These segments were then delivered using the MRLTP dose engine and their dose was warped and summed to the reference grid leading to the respective reconstructed dose. Due to the limited amount of online data, in the cases where treatment time exceeded the available imaging data, the timeline was extended by copying the motion trace (DVFs and dynamic volumes) in an unmirrored fashion.

2.3.1. Adaptive treatment.

2.3.2. Static treatment.

Figures 2(a)–(c) show the mean GTV SI displacement in mm during the simulated treatments from the reference mid-position volume for all subjects. The mean displacement in mm and standard deviation of the motion in the left–right (LR), anterior-posterior (AP) and SI directions was $-0.2 \,(0.3), -0.4 \,(0.1), 3.3 \,(1.1)$ for Patient $1, -0.1\,(0.5), -0.4\,(1.2), -0.5 \,(2.5)$ for Patient 2 and $0.2 \,(0.3), 0.1 \,(0.4), -0.8 \,(1.3)$ for the volunteer respectively. Patient 2 and the healthy subject showed high amplitude breathing cycles with small baseline displacements, with the target being on average at an inferior position compared to the pre-treatment volume. In contrast, patient 1 had a more stable breathing cycle but underwent a baseline shift of approximately 3 mm superiorly after the 4D-MRI and prior to the online imaging.

Zoom In Zoom Out Reset image size

Figure 3 shows the effect of motion on the conventional pre-treatment plans with the 3 mm PTV margin for all subjects. Underdosage can be observed (positive values in the color scales) in the cranial direction in Patient 1 and primarily in caudal direction in Patient 2 and the healthy volunteer. For Patient 1, as expected from the small GTV size and the initial baseline variation, target coverage was greatly reduced leading to a D99% of 23.1 Gy against 26.1 Gy in the REF_PTV plan and a GTV minimum dose of 21.4 Gy compared to 25.7 Gy. While the other two subjects were not greatly affected due to the relatively larger target size and more stable motion, in both cases the minimum GTV dose dropped below the 25 Gy threshold to 24.9 and 24.5 Gy respectively.

Zoom In Zoom Out Reset image size

The intrafraction adapted plans were generated by producing the tightest possible dose distributions around the GTV during the plan adaptations in-between beam deliveries which fulfilled all clinical constraints (table 2). Figure 4 shows the central coronal GTV slice from the ADAPT replanning regime for all subjects.

Zoom In Zoom Out Reset image size

3.3.1. Mid-position update.

Table 3 shows the average GTV displacement relative to the reference imaging, extracted from the updated anatomies used during the inter-beam plan adaptation. These displacements are visualized with horizontal dashed lines in figure 2 (vertical dashed lines indicate the different replanning segments). The mean mid-position GTV displacement was very close to the actual treatment GTV displacement (section 3.1). By comparing the mid-pos displacement of each replanning phase to the true displacement of that period, we establish the performance of the mid-position update regime (table 3, Diff).

Table 3. Mean mid-position displacements (SD) relative to pre-treatment reference volume and differences to the true displacements as calculated by the adaptive replanning pipeline in mm. The two columns show the corresponding values by either including or excluding the first beam targeting the pre-treatment anatomy.

Case	ADAPT Overall	ADAPT (after beam 1)
Pat. 1 Mid-pos	$-0.2, -0.3, 2.8\,(0.1, 0.2, 1.4)$	$-0.2, -0.4, 3.4\,(0.02, 0.02, 0.03)$
Diff	$-0.03, -0.07, 0.56$	$0.002, 0.001, -0.01$
Pat. 2 Mid-pos	$-0.1, -0.3, -0.3\,(0.1, 0.3, 0.4)$	$-0.1, -0.3, -0.4 \,(0.1, 0.2, 0.4)$
Diff	$-0.01, -0.05, -0.06$	$0.001, 0.02, 0.04$
Volun. Mid-pos	$0.1, 0.1, -0.7\,(0.1, 0.1, 0.4)$	$0.2 0.1 -0.8 \,(0.0, 0.1, 0.2)$
Diff	$0.03, 0.03, 0.15$	$0.004, -0.01, -0.04$

Small mean differences are overall observed, only hampered by the baseline variations occurring between the delivery of the first beam and the pre-treatment mid-position. As shown in the last column, the mid-position update employed after the initial beam, greatly decreases these differences, leading to a relative magnitude reduction by 72.2%, 68.9% and 68.3% for each motion component.

3.3.2. Target coverage.

Figures 5(a), (c) and (e) show the DVH plots of the REF_PTV (solid lines), RECON_PTV (dashed lines) and ADAPT (dotted lines) for all subjects.

Zoom In Zoom Out Reset image size

In all cases the online replanning ensured higher GTV coverage than RECON_PTV (25.7 versus 23.1 Gy, 26.1 versus 25.8 Gy and 25.9 versus 25.2 Gy respectively). The minimum GTV dose was maintained above 25 Gy (25.3, 25.2 and 25.2 Gy) while the mean delivered GTV dose was higher (27.3 versus 26.9 Gy, 28.5 versus 27.8 and 28 versus 27.4 Gy).

Regarding baseline variations, the initial 3 mm target shift exhibited by Patient 1 was captured after the first delivered beam leading to full GTV coverage at the end of treatment. This is also visualized in figure 6 where the GTV curve of the motion reconstructed static plan with no PTV margin (RECON), along with the RECON_PTV and ADAPT plans are plotted. The V25Gy (minimum acceptable 99%) increased from 71.5% in RECON to 90.5% in RECON_PTV to 100% in the ADAPT regime.

Zoom In Zoom Out Reset image size

3.3.3. Healthy tissue sparing.

Table 4 presents indicative replanning timings for the different subjects during the individual steps of the adaptive treatment, while using the same number of calculated segments. Each replanning includes the time required to process the input anatomy, delineations and DVFs, calculate the corresponding dose influence data, followed by an iterative ASEQ segment calculation and a final segment weight optimization (SWO). MRLTP replanning ranged from 43.3 s on average for the 6-beam pre-treatment plans to 10.9 s on average for the final single beam plans. For every subject the adaptive replanning took on average 25.6 s, enabling future online usage of MRLTP.

The reference volume was generated by assigning the whole body to water-equivalent tissue (Stam et al 2013), which was then warped accordingly to generate the dynamic volumes. Several techniques (bulk assignment-based, voxel-based and atlas-based) have been developed for MR-only pseudo-CT generation in an effort to fully substitute the electron density information included in a CT scan (Maspero et al 2017). Generally, dose differences in a dynamic setting can be attributed to two factors, the change of the physical path length and the varying tissue density information. The former one forms the main contributing factor in treatment sites with kidney-like tissue properties (Kerkhof et al 2010). Since the path length changes are included in the dynamic volumes produced by the pipeline, we regard the bulk-assignment technique to be adequate for the purpose of this work.

The presented results show the technical feasibility of a replanning framework that utilizes multiple components previously developed in our department. MRLTP has been optimized for the performance requirements of online interventions and the average replanning time of 25.6 s is already promising for online usage in timeframes similar to the presented application. MRLTP was developed to take advantage of parallel CPU/GPU computing and its scalability to higher-end computer configurations should be further investigated.

The other components of the pipeline were originally developed for offline usage and thus should be optimized to qualify for online application. The dose reconstruction module (Glitzner et al 2015) currently performs a beam reconstruction in the order of minutes, while the motion model (Stemkens et al 2016) is able to calculate a new 3D dynamic volume in the order of seconds. We are now working towards the full integration of the individual modules and the optimization of the complete framework.

Due to the high dose 25 Gy single fraction prescription, the simulated treatment resulted in long delivery times of around 18 min. Consequently, the available motion data of 5 and 7 min for the patients and the volunteer respectively had to be concatenated to facilitate the whole treatment time. The range of motion has been correlated to the imaging/treatment times (von Siebenthal et al 2007), which would mean that the actual motion trace could exhibit further drifts and/or amplitude variations as the treatment progresses. Observed over the time scale of a single-beam replanning in this work, these changes could lead either to better agreement between the reference anatomy and the online target position or to larger deviations caused by new baseline variations/drifts. Consequently, these changes would have either neutral or further negative impact to the conventional static treatment. In contrast, our method would lead to tighter distributions in the case of small changes (figures 5(d) and (f)) and ensure target coverage in the case of larger variations (figures 5(b) and 6).

Due to the limited motion data available, the replanning time was not included in the treatment timeline. In this way, the results represent the technical feasibility of fast online adaptation and its possible clinical benefits as replanning times further decrease in the future.

The introduction of intrafraction adaptive treatments will require thorough QA procedures. Both the motion estimation and the online-adapted dose distributions should be carefully evaluated. In Stemkens et al (2016) an MR-compatible motion phantom was used to quantify the geometric accuracy of the model leading to an average difference of 1–1.5 mm. For a simultaneous end-to-end test of both motion estimation and plan adaptation, a motion phantom in combination with 3D dosimetric gel target could be used, ideally with the ability to introduce controlled 3D deformations (Mann et al 2017). However, the relatively high systematic deviations of approximately 5% in gel dosimetry (Vandecasteele and De Deene 2013) should be carefully factored in when evaluating high precision SBRT treatments.

In the proposed framework large anatomical variations during the beam-on time intervals between the online adaptations can produce deviations from the planned dose. This is especially the case for the first beam targeting the reference anatomy and the last beam prior to the end of treatment. These effects can be reduced leading to even better conformality and overall plan quality, by utilizing more beam angles/replanning steps, which would reduce the contribution/dose of each beam and provide a more up-to-date mid-position volume.

The scale of the anatomical changes that could be accounted for by this regime depends on the replanning frequency which is mainly limited by the performance capabilities of the complete system. This includes the individual hardware latencies, imaging and motion estimation as well as the replanning speed. The mid-position update framework allows for a direct correlation between the replanning interval and the underlying anatomical changes. By increasing the replanning frequency, we can progressively account for more subtle changes with the mid-position converging to a single anatomical instance and the replanning to real-time inter-segment plan adaptation.

Under this context the same concept could be applied in dynamic IMRT treatments with volumetric arc therapy (VMAT) where individual beams can be substituted with separate arcs with in-between replanning sections. The replanning frequency will then be correlated with the number and radii of the utilized arcs depending on the application.

Respiratory-gated treatments have been employed in an effort to reduce the free-breathing motion, by delivering radiation only to part of the treatment cycle with the smallest relative residual motion and healthy tissue exposure. The reduced duty cycle can lead to 4–15 times longer IMRT treatments (Keall et al 2006). This increase in treatment time would be restrictive for applications like the single fraction 25 Gy treatment used in this work.

Tracking techniques have also been used to follow the tumor by dynamically adjusting the MLC leaf positions based on several tumor-localization methods, leading to increased target conformality and high duty cycle (Giraud and Houle 2013). However, during tracking the dosimetric impact of the (smaller) out-of-plane motion and, more importantly, the 3D deformations of the target and OARs are not taken into account. Furthermore, during target tracking the spatial dependence of fluence in FFF beams is not addressed.

In our method, we utilize MR-imaging data to extract the 3D anatomical deformations of the target and OARs during treatment and perform inter-beam replanning based on an updated patient volume. The mid-position concept enables high duty cycle treatments suitable for ablative radiotherapy applications. Moreover, the inclusion of the previously delivered dose to the patient into the planning, accounts for the dosimetric effects of residual motion and/or 3D deformations occurring during the beam delivery. Future applications of our method can also be combined with tracking techniques during radiation delivery in-between replannings to compensate for baseline variations occurring during beam-on.