Dosimetric evaluation of an atlas‐based synthetic CT generation approach for MR‐only radiotherapy of pelvis anatomy

Abstract Purpose To investigate the potential of an atlas‐based approach in generation of synthetic CT for pelvis anatomy. Methods Twenty‐three matched pairs of computed tomography (CT) and magnetic resonance imaging (MRI) scans were selected from a pool of prostate cancer patients. All MR scans were preprocessed to reduce scanner‐ and patient‐induced intensity inhomogeneities and to standardize their intensity histograms. Ten (training dataset) of 23 pairs were then utilized to construct the coregistered CT‐MR atlas. The synthetic CT for a new patient is generated by appropriately weighting the deformed atlas of CT‐MR onto the new patient MRI. The training dataset was used as an atlas to generate the synthetic CT for the rest of the patients (test dataset). The mean absolute error (MAE) between the deformed planning CT and synthetic CT was computed over the entire CT image, bone, fat, and muscle tissues. The original treatment plans were also recomputed on the new synthetic CTs and dose–volume histogram metrics were compared. The results were compared with a commercially available synthetic CT Software (MRCAT) that is routinely used in our clinic. Results MAE errors (±SD) between the deformed planning CT and our proposed synthetic CTs in the test dataset were 47 ± 5, 116 ± 12, 36 ± 6, and 47 ± 5 HU for the entire image, bone, fat, and muscle tissues respectively. The MAEs were 65 ± 5, 172 ± 9, 43 ± 7, and 42 ± 4 HU for the corresponding tissues in MRCAT CT. The dosimetric comparison showed consistent results for all plans using our synthetic CT, deformed planning CT and MRCAT CT. Conclusion We investigated the potential of a multiatlas approach to generate synthetic CT images for the pelvis. Our results demonstrate excellent results in terms of HU value assignment compared to the original CT and dosimetric consistency.


| INTRODUCTION
The use of magnetic resonance imaging (MRI) for radiotherapy application has been rapidly increasing in recent years. 1 The main advantage of MRI is superior soft tissue contrast that improves the delineation of target volumes and organs at risk (OARs). Despite this clear superiority for tissue contouring, there are concerns regarding errors introduced by mis-registration between the diagnostic MRI and radiotherapy planning CT or differences in bladder and rectum filling even if the MRIs are acquired in radiotherapy position. The idea of making MRI as the primary image set for radiotherapy planning and synthesizing a CT from the MRI information eliminates this concern and has enabled MR-only radiotherapy approaches.
Various methods for generating synthetic CT images for pelvic anatomy have been introduced in the literature. [2][3][4][5][6][7][8][9][10] Among all these promising approaches, MRCAT (MR for Calculating ATtenuation) available on 3T Philips Ingenia platform 11 is one of the few commercial products 11,12 being used in our clinic for MR-only radiation therapy. 13 MRCAT CT is generated from a 3D mDixon fast field dual echo sequence by creation of three distinct images: water only, fat only, and in-phase MRI. These image series are utilized in a classification algorithm to provide soft tissue and bony clusters. These two clusters are further divided into water, adipose, cortical, and spongy bones. Each class of tissue is then assigned a bulk electron density.
Although, MRCAT has been successfully applied in the clinic, 13 the algorithm is currently limited to Philips MR scanners only. An ideal synthetic CT generation method would be independent of MR vendor and/or MR sequence. The MRCAT algorithm is also currently limited to generate bones till L4 which is not ideal if there is an intent to treat nodes higher than L4 for some prostate cases. Hence, alternative, more generally applicable methods for synthetic CT generation are still needed.
In this study, we aim to investigate the potential of a multi atlasbased approach originally developed for head and neck anatomy 14 to generate synthetic CT images for patients undergoing radiotherapy for prostate cancer. Several steps in the original algorithm were modified to expand its use to pelvic anatomy. We compare the image characteristics and dosimetric results to those of the deformed planning CT as well as MRCAT CT.

2.A | Image acquisition
After obtaining IRB approval, 23 sets of CT and MR images were retrospectively selected from a pool of prostate cancer patients (aged 54-87) for whom mDixon-based MRCAT CT scans were also available. No prior assumption was made in terms of image quality to select this patient cohort. All patients received radiation therapy in our institution with a prescription dose ranging from 25 to 72 Gy using either five fraction stereotactic body radiosurgery or conventionally fractionated intensity modulated radiotherapy. For seven patients, the external beam radiotherapy (25 Gy in five fractions) was administered following brachytherapy. All CT and MRI scans were acquired in the treatment position. CT scans (either Philips Healthcare, Cleveland, OH (n = 21) or GE Medical System, Chicago, IL, USA) were acquired in the helical mode with a tube voltage of 120 kV, slice thickness of 2.5 mm, matrix size of 512 × 512, and in-plane pixel size of 0.9766 × 0.9766 mm 2 . MR scans were acquired on a Philips 3T (Philips Healthcare, Cleveland, OH) Ingenia system using a vender-provided phasedarrayed anterior and posterior coils, and included a dual fast field echo mDixon (in-phase, out-phase, fat, and water) sequence with TE1/TE2/ TR = 3.3/4.6/6.07 ms, flip angle = 10 o , slice thickness of 2.5 mm and in-plane pixel size of~1 mm 2 . All MR scans were acquired with 30 cm length in superior-inferior direction limited superiorly to L4. Ten of 23 patients were randomly selected to create the atlas and the remaining patients were reserved for the test dataset. MR scans contained noticeable artifact in the most inferior and superior slices. These slices were removed from our atlas resulting to lack of data near those regions.

2.B | Image preprocessing
All MR scans were automatically preprocessed in two steps prior to synthetic CT generation. In the first step, an image analysis technique 15 was utilized to reduce the intensity inhomogeneity due to field nonuniformity, tissue susceptibility effects, and scanner-dependent variabilities. Local clustering properties of the image intensities were extracted using a model of intensity inhomogeneity surrounding each pixel to estimate the regional signal loss due to bias fields inhomogeneity. The original image was then corrected accordingly. This procedure was applied along the sagittal direction since this is the direction of more pronounced field inhomogeneity. 14 In the next step, a landmark-based standardization technique was used to standardize the MR intensity histogram.
This reduced the scanner-dependent variation in MR image intensities and facilitates the registration process. We applied the above procedure to water-and fat-only images (Fig. 1). To find the tissue-specific landmarks, fuzzy c-means clustering was initially applied to the water-and fat-only images to classify each image into dark and bright regions. Bright regions represent the muscle and fat tissues in the water-and fat-only images respectively. The where in the above, MR S,W and MR S,F represent the standardized water-and fat-only images and MR W,FE denotes the standardized fatenhanced water-only images [ Fig. 1(c)]. α = 0.5 was used in this paper.

2.C | CT-MR atlas
The CT number-suppression approach developed in our previous synthetic CT method 14  During deformation, a subsampling rate of 1 × 1 × 1 was used to avoid smoothing and blurring effects. Mean square error was utilized as the cost function to fine-tune the rigidly aligned images. The resulting deformation matrix was then applied to the original planning CT to obtain the deformed planning CT-MR pairs (CT reg , MR W, FE ) that form the atlas. It is worthwhile to note that the purpose of synthetic CT generation is to assign CT number to each MR voxel.
Hence, MR geometry is our ground truth and we should deform everything onto the MR images.

2.E | Evaluation of results
Two methods were used to evaluate the performance of the proposed approach to generate synthetic CTs of the pelvic anatomy. In the first method, a dataset comprised of the ten patients whose Example of water-only (a), fatonly (b), and fat-enhanced water-only image (c). As shown in (c), the contrast between fat and air regions was improved in fat-enhanced water-only image compared to water-only image. Water-only image was used to create the CT-MR atlas and fat-enhanced water-only image was used for atlas propagation.
CT-MR pairs were used for atlas creation (training dataset) was used to generate synthetic CT for each patient in a leave-one-out scheme through which we also incorporated the training dataset into evaluation as well. Each patient was sequentially considered as a "new" patient, a CT-MR atlas was formed from the remaining nine patients, and a synthetic CT was generated using the "new" patient wateronly MR image. In the second method, synthetic CTs were generated for the remaining 13 of the original 23 patients (test dataset) as an independent validation. For these patients, the CT-MR atlas generated from the first ten patients was applied. As mentioned, for all 23 patients, the mDixon-based MRCAT CTs were also available.
To quantify the voxel-level accuracy of the intensities for each synthetic CT generated for evaluation, the mean absolute error (MAE) between the synthetic CT and its corresponding deformed planning CT (CT reg ) and MRCAT CT scan, was computed over the entire image, as well as the bone, fat, and muscle regions separately.
To identify the evaluation region for the entire image, a mask was applied to exclude the background from analysis. To obtain the bone, fat, and muscle regions, the corresponding clusters from the MRCAT CT images were identified and applied to the synthetic CT and CT reg images. MRCAT has the same geometry as MRI and has an excellent classification result on fat and muscle. Cortical and spongy clusters were lumped together to produce the bony areas.
To evaluate the suitability of the synthetic CT for radiotherapy dose calculation, the patient's treatment plan, originally generated on the planning CT, was transferred to the deformed planning CT   For comparison purpose, the dose difference between the deformed planning CT and plan with no inhomogeneity correction were also included for the selected structures. As noted, the largest dose difference in the synthetic CT was less than 2.5% and seen in small bowel.
Our further investigation revealed that this dose difference is mainly

| DISCUSSION
In this work, we modified and extended a previously presented multiatlas approach 14 to generate synthetic CT for pelvic anatomy. In this modified version, water-only MRI, rather than in-phase MRI, was To evaluate the performance of the proposed methodology, the patient image sets were separated into a training dataset used for the atlas creation and a test dataset used for independent validation of the method. Synthetic CTs for the patients in the training set were generated using a leave-one-out approach. The synthetic CTs generated for both datasets were compared to MRCAT CT, a T A B L E 1 The average of planning target volume (PVT) and various organs at risk (%) dose difference between the plan calculated using the deformed planning CT (CT reg ) and the plans calculated using the synthetic CT or MRCAT CT for patients in the training dataset. NIC: No inhomogeneity correction. hour for an image volume with a size of 512 × 512 × 120 pixels.
Therefore, even with GPU programming, the entire process may require several hours to generate a single synthetic CT. Expediting the registration process and GRE calculation are important areas for future study. In addition, as shown in Fig. 3, MRCAT (and bulk density assignment approaches, in general) produce very sharp and clean images while the proposed atlas-based approach generates a more blurred image. This may produce some difficulties if the synthetic CT is used as the primary image set for contouring of certain structures especially bony regions like femoral heads which are easier to contour on CT. Incorporating additional information from fat-only and in-phase images into generation of the synthetic CT may ultimately yield sharper images and is also an area for further investigation.
Furthermore, the use of multiparametric GRE calculation is also part of our future study. Currently, we calculate the generalized registration error using the difference map between the two coregistered fat-enhanced water-only images. Applying the deformation matrices to all standardized MR images, including in-phase, fat-only, and water-only image series, and utilizing them for similarity measurement, presumably provides more information to find the best match of a voxel in a new patient among the ones in the atlas.

| CONCLUSION S
We have modified and extended a previously described multiatlas approach focused on the head and neck region to generate synthetic CTs for pelvic anatomy. The results were compared to MRCAT CT, a commercially available product for radiotherapy use.

ACKNOWLEDG MENT
This research was funded in part through NIH/NCI Cancer Center support grant/core grant P30 CA008748.

CONFLI CT OF INTEREST
This work was partially supported through a Research Agreement from Philips Healthcare, Cleveland, OH.