Evaluation of a commercial synthetic computed tomography generation solution for magnetic resonance imaging‐only radiotherapy

Abstract Purpose To evaluate the Siemens solution generating Synthetic computed tomography (sCT) for magnetic resonance imaging (MRI)‐only radiotherapy (RT). Method A retrospective study was conducted on 47 patients treated with external beam RT for brain or prostate cancer who underwent both MRI and CT for treatment planning. sCT images were generated from MRI using automatic bulk densities segmentation. The geometric accuracy of the sCT was assessed by comparing the Hounsfield Units (HU) difference between sCT and CT for bone structures, soft‐tissue, and full body contour. VMAT plans were computed on the CT for treatment preparation and then copied and recalculated with the same monitor units on the sCT using the AcurosXB algorithm. A 1%‐1mm gamma analysis was performed and DVH metrics for the Planning Target Volume (PTV) like the Dmean and the D98% were compared. In addition, we evaluate the usability of sCT for daily position verification with cone beam computed tomography (CBCT) for 14 prostate patients by comparing sCT/CBCT registration results to CT/CBCT. Results Mean HU differences were small except for the skull (207 HU) and right femoral head of four patients where significant aberrations were found. The mean gamma pass rate was 73.2% for the brain and 84.7% for the prostate and Dmean were smaller than 0.5%. Large differences for the D98% of the prostate group could be correlated to low Dice index of the PTV. The mean difference of translations and rotations were inferior to 3.5 mm and 0.2° in all directions with a major difference in the anterior‐posterior direction. Conclusion The performances of the software were shown to be similar to other sCT generation algorithms in terms of HU difference, dose comparison and daily image localization.


| INTRODUCTION
The main reason for including magnetic resonance imaging (MRI) in the Radiation Therapy (RT) workflow is the higher soft tissue contrast compared to CT. MRI cannot be used directly for dose calculation because MR intensities correlate with proton densities and relaxation properties whereas dose calculation in treatment planning systems requires data on electron density from CT. 1 Therefore, a combined workflow is used including both MRI for tissue segmentation and CT for dose calculation after MRI-CT registration, which can introduce uncertainties estimated to be up to 2-5 mm. 2 Interest is growing to use MRI as the only modality in RT to take advantage of its soft tissue contrast and remove inter-modality registration uncertainties. 3 Recent studies have investigated the feasibility of implementing an MR-only workflow by synthetically generating CT images, called synthetic-CT (sCT), directly obtained from MRI to enable dose calculation and position verification. 4 Recently, SIEMENS HEALTHINEERS ® commercialized a solution which generates sCT using a bulk density method. In this method, MRI voxel intensity information is translated to CT numbers and sCT images. 5 Bulk density techniques either used a single homogeneous or a multiple tissue override, which improve the dose calculation but required long manual contouring. 6 The imaging software provided by SIEMENS HEALTHINEERS ® generates the sCT image using an automatic tissue classification method with five classes. To the best of our knowledge, there is no study evaluating the performances of this solution. The objective of the study was to evaluate the geometric accuracy, dose calculation and positioning performance of this commercial sCT approach in external beam radiotherapy for brain and prostate cancer.

2.A | Patient data, image acquisition, and sCT generation
Retrospective analysis was performed using MRI and CT data from 47 patients (16 primary brain tumors and 31 prostate cancer).
Approval for the study protocol was obtained from the medical research ethics committee and informed consent was obtained from all patients. See Table 1 for more information about RT volumes and prescribed doses.
Patients first underwent a CT acquisition with a SIEMENS SOMATOM Confidence RT scan. Then, MR images were acquired using a 1,5 T SIEMENS MAGNETOM AERA XJ MRI scan ® with a mean time of 4 days (range 0-11) between the two imaging sessions. MRI sequences were acquired to generate sCT images in addition to routine sequences used for delineation (see Supplemental   Table S1 and S2 for magnetic parameters details). The CT and MR images were acquired in the treatment position before RT. For brain imaging, the head was immobilized in a thermoplastic three-point mask during both planning CT and MRI. Masks were marked at the CT session and patient repositioning at the MRI session was checked with a laser system. For prostate imaging, patients were positioned in a supine position on the provided flat table, with identical immobilization thanks to a knee support cushion. Patients were tattooed at T A B L E 1 Volumes, prescribed doses and type of RT for prostate and brain cancer groups.

2.B | Structures delineation
The study and standard workflows are described in Fig. 2

2.C | Geometric accuracy
A comparison was made between the HUs of the sCT and the HUs of the reference CT scan for bone structures, soft-tissue and body contour in prostate and brain groups. Bone segments were generated by thresholding the respective images at 100 HU, followed by a morphological hole filling. Soft tissue segments were generated by thresholding the sCTs and CTs at −100 HU and subtracting the previously generated bone structures. Body contour segments were similarly generated by thresholding at −350 HU, followed by a morphological hole-filling. Mean HU were extracted in each volume and differences were calculated between the CT and sCT for each patient.   was calculated using either the CT or the sCT as reference and mean differences were calculated using the following equation:

2.E | Positioning performance
where ΔV CT=sCT is the intrinsic offset between CT and sCT.

| RESULTS
Three prostate patients were excluded from the study (2 having an artefact or a prothesis on MRI and 1 because MRI did not encompass the entire body).

3.A | Geometric accuracy
Average HU value and mean differences between CT and sCT are given in Table 2 Figure S1).
Except those patients, the mean HU difference for femoral heads is improved (19.8 AE 10.7 HU).

3.B | Dose comparison
Results of the gamma pass rate and mean dose difference analysis are plotted in Fig. 3

| DISCUSSION
For the brain, mean HU differences were small, except for the bones. The high HU of the skull is known to be underestimated by sCT. 7 For the prostate, mean HU differences observed for the entire pelvis were greater than those reported in the literature. 8,9 Significant aberrations exist on sCT images reconstruction, currently under investigation by SIEMENS ® .
The dosimetric comparison of plans optimized on the CT and calculated on the CT and sCT for the brain showed similar results in the PTV region, except when air cavities were in close proximity.
For the prostate group, large differences were observed for the D 95% and D 98% . However, the dose comparison conducted encompasses both differences linked to the change of modality (CT/sCT) and to the change of PTV contours (CT and MRI respectively). The DICE index of two volumes is defined as the intersection of the volumes divided by the union of the volumes. 17 Figure 6 shows the D 98% metric observed for the prostate plans as a function of the

| CONCLUSION
This is the first study evaluating the sCT generation method proposed by SIEMENS ® for brain and prostate locations using a high sensibility algorithm for the dosimetric analysis (AcurosXB). The performances of the software were evaluated in terms of HU difference, dose comparison and daily image localization and showed reasonable deviations between CT and sCT. The largest differences of the dose comparison could be related to patient repositioning between the CT and MRI.

ACKNOWLEDG MENTS
The authors thank Florence Legouté, Thibaut Lizée and Nathalie

CONFLICT OF INTEREST
None.

SUPPORTING INFORMATION
Additional supporting information may be found online in the Supporting Information section at the end of the article.  and additional sequences needed for s-CT generation (blue) and additional sequences needed for s-CT generation (blue) GONZALEZ-MOYA ET AL.