Study population
Patients receiving curative radiotherapy for prostate cancer at our institute between October 2018 and May 2019 were asked to participate in this study. In accordance with the regulations of the local ethics committee, 31 patients gave their informed consent to add an MRI sequence for generating a sCT to the standard clinical MRI acquisitions used for target delineation.
Patients with a 3 T MRI contra indication were excluded, as well as patients with hip prostheses or with an abdominal diameter exceeding 50 cm in left-right or 30 cm in anterior-posterior direction, since a sCT could not be generated for these cases.
CT/MR workflow
Following the current clinical workflow, four cylindrical gold FMs (diameter 1 mm, length 5 mm, RT-Idea B.V.) were implanted in the patient’s prostate, at least four days prior to CT- and MR imaging. To minimize anatomical changes between the scans, the time between CT and MR scans was less than two hours.
The planning CT was acquired with an in-plane resolution of 1 × 1 mm² and a 2.5 mm slice thickness (LightSpeed RT16 CT, GE). The patients were instructed to have a filled bladder for both the CT examination and each radiotherapy fraction, which makes it possible to spare the bladder during treatment. To achieve this, patients were asked to empty their bladder and to subsequently drink 500 ml of water 1.5 hours before the appointment.
The MRI was acquired in treatment position (Ingenia 3T MRI with RT Oncology configuration, Philips Healthcare, Best, The Netherlands). The patients were not instructed to have a filled bladder for the MRI examination and were allowed to urinate before the MRI examination, if necessary. The standard clinical acquisition protocol consisted of transversal and sagittal T2-weighted images (which are referred to as T2WTRA and T2WSAG) for prostate delineation, covering the entire prostate and seminal vesicles (slice thickness 3 mm, in-plane resolution 0.6 × 0.7 mm², see Supplemental Materials, Table A1), and a balanced turbo field echo acquisition with fat suppression, i.e. spectral attenuated inversion recovery (referred to as FMimage) to visualize the FMs (slice thickness 1 mm, in-plane resolution 1.0 × 1.0 mm²).
The FMimage was used to facilitate a FM-based registration of the T2WTRA to the CT, because FM were hardly visible on the T2WTRA (Fig. 1). First, the T2WTRA was automatically rigidly registered to the FMimage using mutual information as metric and a rectangular region of interest (ROI) containing the prostate but no bony structures (labelled FMimage-T2WTRA). Then, the FMimage was manually registered to the CT, by aligning the FMs (labelled FMimage-CT). Lastly, the CT and the TW2TRA scans were registered by performing the former two registrations in succession (labelled CT-T2WTRA, see Fig. 1). All registrations were performed in Velocity, R4, Varian Medical Systems.
A radiation oncologist contoured the CTV (consisting of the prostate and base of the seminal vesicles) on the T2WTRA, which was fused with the CT using the CT-T2WTRA registration, according to the EAU-ESTRO guidelines for prostate cancer (28). For some patients, elective lymph node regions were also included and contoured by the radiation oncologist on CT images. Specialized radiation therapists (RTTs) contoured the OAR (rectum, anal canal, bladder and both hips, as well as the sigmoid, small bowel and bowel bag in proximity of the PTV) on CT images according to department’s protocol.
Different dose prescriptions were used per patient, depending on the classification of the prostate cancer. The treatment planning technique consisted of 10 MV dual arc volumetric modulated arc therapy (VMAT) (Raystation, R7, RaySearch Laboratories).
The IGRT was based on daily online registration using CBCT (slice thickness 0.5 mm, in-plane resolution 0.5 × 0.5 mm²). An automatic bone match was performed to evaluate the patient set-up with a ROI containing the small pelvis, followed by an automatic FM match (XVI, R5.0.4, Elekta).
Simulated MR-only workflow
The MR-only workflow was simulated by replacing the CT with the sCT. A schematic overview of both the CT/MR- and the simulated MR-only workflow is shown in Fig. 2.
To generate the sCT an additional mDIXON fast field echo (FFE) MRI sequence (slice thickness 2.5 mm, in-plane resolution 1.7 × 1.7 mm², other clinical MRI scan acquisition parameters can be found in Supplemental Materials, Table A1) was acquired, sliced in the transverse direction (MRCAT, RTgo plugin 3.0, Philips, 2016). From the mDIXON scan the in-phase, water and fat reconstructions were used to generate the sCT. FOV of the image was from L4/L5 to the caudal border of the symphysis, covering the entire body contour in AP and RL directions. MRCAT is an algorithm using the mDIXON and a model based segmentation of the bones to create a sCT that consists of five different materials: air, compact and spongy bone, fat and water-rich tissue (3). The mDIXON and the T2WTRA were acquired in succession to minimize organ motion between the scans. From the mDIXON acquisition a water-only image (mDIXON-w) was reconstructed, on which the FM were visible as signal voids (see Fig. 3A). The T2WTRA was registered to the mDIXON-w using an automated registration based on gray values with mutual information as metric and a rectangular ROI containing the prostate without including bony structures. If necessary, the automatic registration was manually adjusted based on the signal voids of the FM on the MRI. Because the sCT was generated from the mDIXON, the T2WTRA image and the sCT registration was identical to the T2WTRA to mDIXON-w registration (labelled sCT-T2WTRA), shown in Fig. 1.
The FMs are not visible on the sCT scan, since it identifies only five materials. To automatically identify the FMs on the mDIXON-w scan is challenging without any prior input on the positions of the FM, potentially giving rise to wrongly identification of features showing up as a signal void (e.g., calcifications) (29). Therefore, we developed a semi-automatic method to identify and burn-in the FM on the sCT. In the first manual step, a dedicated MRI sequence (FMimage) was registered to the mDIXON-w scan to help identify FM positions on the mDIXON-w scan. Once identified, the signal void associated with a FM was manually delineated on three slices of the mDIXON-w scan and isotropically expanded by 2 mm. In the second automatic step, the six voxels with the lowest intensity within the ROI were determined and were assigned a CT number of 3000 Hounsfield Units (HU). Using an in-house developed C + + software tool, the FMs were burned-in on the sCT (Fig. 3A -3E).
To enable image registration of the sCT with CBCT for position verification a copy RT plan was created on the sCT (RayStation). Image registration was performed in the same way as for the CT/MR workflow (XVI, Elekta) (Fig. 2).
To compare the accuracy of the CT/MR- and the MR-only workflow three sources of inaccuracy were investigated:
-
the inter-observer variation of the image registration for target delineation (CT-T2WTRA versus sCT-T2WTRA),
-
the inter-observer variation of burned-in FM positions using the semi-automatic method (only in MR-only workflow),
-
the accuracy of FM-based image registration for position verification (CT to CBCT versus sCT to CBCT).
The inter-observer variation of image registration for target delineation
Seven experienced observers (five RTTs and two medical physicists) performed image registration for target delineation for both the CT/MR workflow (CT-T2WTRA) and the MR-only workflow (sCT-T2WTRA) for the first twenty patients consecutively included in this study. The resulting registrations, which were labelled CT-T2WTRA for the CT/MR and sCT-T2WTRA for the MR-only workflow (Fig. 1), were specified by six parameters: the translation in right-left (RL), cranio-caudal (CC), and anterior-posterior (AP) direction, and the rotation angle about the RL, CC, and AP axis. For each parameter the inter-observer registration error (IOE) was quantified as the variation of registration results. For the individual patients, the inter-quartile range (IQR) was calculated. To be able to pool the data between patients, the mean of the registration values per patient was subtracted. For the cohort of patients, the IOE of the registration parameters was described by the standard deviation (SD).
A paired two-sided Wilcoxon signed-rank test with a significance level of 5% was used to test the difference in IQR. A non-parametric Levene’s test with a significance level of 5% was used to test the difference in SDs for the registration results of the pooled data, e.g. for all observers and patients combined (data analyzed with IBM SPSS Statistics for Windows, Version 24.0 (NY: IBM Corp)).
The inter-observer variation of burned-in FM positons using the semi-automatic method
The seven observers (five RTTs and two medical physicists) delineated the signal voids associated with the FM on the mDIXON-w image for the first twenty patients consecutively included in this study It was determined whether all observers were able to agree in correctly identifying the same FMs. Then the center of mass (COM) of the 1, 2, 3 up to 15 voxels with the lowest intensity within each delineation was calculated.
To quantify the inter-observer variation for each marker the SD of the RL, CC and AP component of the COM position was calculated. The SD of the FM positions for 1–15 burned-in voxels were compared.
The accuracy of FM-based image registration for position verification
All available CBCT scans were included of patients in this study for whom the generation of the sCT was successful. As described earlier, signal voids caused by the FM were delineated and for each FM six voxels were burned-in on the sCT. These delineations were created by an experienced RTT and checked by a medical physicist. The sCT with burned-in FM was retrospectively registered with the CBCT by an experienced RTT in the same way as for the CT/MR workflow. The CT/MR workflow was considered the gold standard for image registration for position verification.
For a proper comparison of sCT-CBCT and CT-CBCT registrations, the different patient position and anatomy on sCT and CT needed to be taken into account. This was achieved by correcting the sCT-CBCT registrations using a FM-based registration between CT and sCT using an in-house developed software tool (Supplemental Materials). Remaining discrepancies between FM positions on sCT and CT after this rigid registration were caused by prostate deformation and inaccuracies in determining center of mass (COM) positions of the FMs on the different modalities, and were quantified by calculating the root mean square (RMS) of the components of the difference vector of the COM positions of all markers of all patients.
After this correction was applied for each patient, the average and SD of the translation in RL, CC, and AP direction, and the rotation angle about the RL, CC, and AP axis of the available CBCT registrations were calculated for both CT/MR and MR-only workflow. These were used to determine population mean, systematic, and random error per workflow. The distributions for the CT/MR and MR-only workflow were compared by using a non-parametric Wilcoxon signed-rank test with a significance level of 5%, to evaluate whether the difference in distributions between both workflows was statistically significant (SPSS).
Additionally, for each CBCT the difference between the sCT-CBCT and CT-CBCT registration was determined. For each patient the average and SD of the difference in translation in RL, CC and AP direction, and rotation angle about the RL, CC and AP were calculated for the available CBCT registrations of this patient. These were used to determine population mean, systematic and random errors. A non-parametric Wilcoxon signed-rank test with a significance level of 5% was used to test whether population mean was different from zero (SPSS).