MRI‐based synthetic CT shows equivalence to conventional CT for the morphological assessment of the hip joint

Abstract This study evaluated the accuracy of synthetic computed tomography (sCT), as compared to CT, for the 3D assessment of the hip morphology. Thirty male patients with asymptomatic hips, referred for magnetic resonance (MR) imaging and CT, were included in this retrospective study. sCT images were generated from three‐dimensional radiofrequency‐spoiled T1‐weighted multi‐echo gradient‐echo MR images using a commercially available deep learning‐enabled software and were compared with CT images through mean error and surface distance computation and by means of eight clinical morphometric parameters relevant for hip care. Parameters included center‐edge angle (CEA), sharp angle, acetabular index, extrusion index, femoral head center‐to‐midline distance, acetabular version (AV), and anterior and posterior acetabular sector angles. They were measured by two senior orthopedic surgeons and a radiologist in‐training on CT and sCT images. The reliability and agreement of CT‐ and sCT‐based measurements were assessed using intraclass correlation coefficients (ICCs) for absolute agreement, Bland–Altman plots, and two one‐sided tests for equivalence. The surface distance between CT‐ and sCT‐based bone models were on average submillimeter. CT‐ and sCT‐based measurements showed moderate to excellent interobserver and intraobserver correlation (0.56 < ICC < 0.99). In particular, the inter/intraobserver agreements were good for AV (ICC > 0.75). For CEA, the intraobserver agreement was good (ICC > 0.75) and the interobserver agreement was moderate (ICC > 0.69). Limits of agreements were similar between intraobserver CT and intermodal measurements. All measurements were found statistically equivalent, with average intermodal differences within the intraobserver limits of agreement. In conclusion, sCT and CT were equivalent for the assessment of the hip joint bone morphology.


| INTRODUCTION
The initial diagnosis and evaluation of hip structural disorders, such as hip dysplasia or femoral acetabular impingement, are generally performed on anteroposterior and lateral radiographs. However, because radiographs only represent a two-dimensional (2D) projection, they might not reflect the full 3D variation in bone shape resulting from the disorder. 1 3D imaging techniques provide a visualization of the entire hip anatomy and enable postacquisition 3D reformatting to standardize patient positioning, 2 as patient positioning might affect the diagnosis. 3 As a result, 3D imaging, whether based on magnetic resonance imaging (MRI) or on computed tomography (CT), has been shown to improve the diagnosis 4,5 and the surgical planning 6,7 of hip morphological disorders. In addition, MR and CT have similar diagnostic power, providing accurate bone models 8,9 and morphometric measurements of the hip which correlate well with each other 10,11 and with radiography-based measurements. 12,13 MR images are commonly used for diagnostic purposes in orthopedic care 14,15 due to their ability to expose defects in periarticular and intraarticular soft tissues. 16 However, the nonselective visualization of bone on common MR images complicates both bone modeling and the measurement of diagnostic parameters as extra care needs to be taken to discriminate bone from soft tissues such as the labrum 13 or ligaments. 17 CT has traditionally been the modality of choice for the assessment of osseous structures, enabling 3D bone visualization for diagnostic purposes 18 and for a range of motion analysis 5,19 with bone models generated faster than with MR images. 20 However, CT imaging introduces an adverse radiation burden, 11 especially for younger populations. Low-dose CT techniques have been developed in the last decade to limit the radiation burden 21 but when bone and soft tissue information is required, two modalities still have to be acquired and processed.
To produce a radiation-free alternative that would provide accurate morphometric measurements for diagnosis whilst enabling fast and accurate bone modeling for planning, CT surrogates could be obtained from MR. Such a unimodal workflow would reduce patients' burden and simplify clinical workflow. Accordingly, MR sequences have been developed to acquire images with CT-like contrast, of which the most promising is zero-echo time (ZTE) imaging. 10,22 However, this technique is not quantitative, requires dedicated hardware, and is prone to false-positive bone identification at water-fat interfaces and fascia. 22 Alternatively, MR-based synthetic computed tomography (sCT) offers a quantitative CT-like contrast, intrinsically registered to the MR images. Although thoroughly investigated for radiotherapy treatment planning and positron emission tomography-MRI attenuation correction, 23 the use of sCT for orthopedic purposes is limited. Recent studies reported promising results, demonstrating overall accurate bone geometry on sCT in lower arms in an ex vivo setting, 24 and in vivo in the cervical spine, 25 lumbar spine, 26 and in the sacroiliac joint. 27 The aim of this study was to evaluate the accuracy of sCT, as compared to CT, for the 3D assessment of the hip morphology. We compared the morphology of the hip joint as assessed on CT and sCT using global surface distance metrics and local morphometric parameters that are clinically relevant for diagnostic indications in orthopedic care. It was hypothesized that bone morphology and contrast are reconstructed accurately by sCT generation models, thus providing a radiation-free time effective method for diagnostic and planning in hip care.

| METHODS
This retrospective equivalence study was performed in accordance with the regulations of the local medical ethical committee, and waiver of written informed consent was obtained (18-381/C).

| Data collection
Imaging datasets of male patients were randomly collected from an existing radiotherapy database containing patients who underwent CT and MRI between October 2017 and April 2018 for the treatment of prostate cancer. Only patients without any implants were included.
MR images were acquired using a 3T scanner (Ingenia; Philips Healthcare), using a torso coil in combination with a multi-echo gradient-echo sequence. Acquisition parameters included echo times of 2.1, 3.5, and 4.8 ms, a repetition time of 6.5 ms, a total acquisition time of 2 min 38 s, and a flip angle of 10°. Images were acquired axially at a resolution of 1.2 mm × 1.2 mm × 2 mm and were reconstructed from the k-space by the scanner at a resolution of 0.97 mm × 0.97 mm × 1 mm, in a 448 × 448 × 160 matrix.
CT scans (Brilliance CT Big Bore; Philips Healthcare) were reconstructed at a slice spacing of 3 mm and a pixel spacing ranging from 0.8 to 1.1 mm as per the standard radiotherapy clinical protocol. MR and CT images have been acquired within 1 h, in head-first supine position.
sCT images were generated fully automatically from the first two MR echoes using a deep learning-enabled software for sCT generation (BoneMRI v1.1; MRIguidance B.V.). The software is based on a 3D patch-based UNet-like neural network 28,29 that was trained on patients from a similar cohort (radiotherapy patients). Images thus generated have the same resolution, orientation, and matrix size as the MR images. sCT images were generated in 2 min 53 s on a GeForce RTX 2080 Ti (NVIDIA) graphics processing unit.

| Bone morphology and contrast
Bone morphology and contrast on sCT images were validated against CT by means of mean error and surface distance metrics. Mean error expresses the voxel-wise difference between CT and sCT and reflects the difference in contrasts between both modalities. Surface distance measures the distance for each vertex on a CT-based bone FLORKOW ET AL.
| 955 model to the closest point on the sCT-based bone model and vice versa (sCT to CT). The root-mean-square error (RMSE) of the surface distance was computed as an overall indication of the morphological differences between bone structures in CT and sCT. To compute these metrics, bones were semi-automatically segmented on CT and sCT images. The segmentation was initialized with in-house deep learning software, extensively manually edited using 3D Slicer 30 and manually checked by a second observer. Then CT and sCT images were rigidly registered using the Elastix registration toolbox. 31 The registration process applied an Euler transform on the bones to minimize the intermodal advanced Mattes mutual information using adaptive stochastic gradient descent. 31 The registration was done independently for the femoral and pelvic bones.

| Hip joint morphometric parameters
The local geometry of the hip joint as visualized on sCT images was validated by means of eight morphometric parameters that were measured by visual annotation on CT and sCT images.   Figures 1B and 1D. These parameters were measured as they are used in the management and preoperative assessment of orthopedic disorders. [33][34][35][36] To extract the aforementioned parameters, anatomical landmarks were annotated by three readers on the images as presented in Figures 1A and 1C. The desired distances and angles were subsequently automatically computed using Matlab 2017a (Math-Works, Inc.) using the coordinates of the annotations.

• Readers
Two senior orthopedic surgeons (R.S. and B.W., with a specialist experience of 23 and 12 years, respectively) and a radiologist intraining with a specialization in musculoskeletal radiology (W.F.) independently identified the anatomical landmarks on the images.
Readers annotated the landmarks independently and were blinded to the other readers' measurements. CT and sCT were randomly shuffled for the annotations and no mention was given to whether a CT or sCT was being annotated. For the assessment of the intraobserver variability, R.S. repeated his annotations with a 1-month interval.

| Statistical analysis
Reliability was measured by means of intraclass correlation coefficients (ICCs) for absolute agreement for the inter-and intraobserver variabilities. 37 The CT-to-sCT intermodal agreement was assessed using a Bland-Altman analysis. 38 The equivalence between CT and sCT was tested for each measurement using paired two one-sided tests (TOST). 39 This test checked whether the average difference between the CT-and sCT-based measures differed by more than a user-defined equivalency margins (±Δ). Δ was defined as the intraobserver limit of agreement (LoA), computed as 1.96*σ intra , where σ intra is the intraobserver standard deviation obtained from the literature. When σ intra was not available, the standard deviation of the interobserver variability, σ inter , was used instead. Values for the reference inter-and intraobserver LoAs are given in Figures 1B and 1D. 10,36,40,41 TOSTs were performed separately on the left and right hips to meet the data independence assumption required by the statistical test. A Bonferroni correction was applied to correct for the 16 repeated comparisons (8 parameters, left/right for data independence). As such, p < 1.6E−3 was considered significant. The normality of the data was determined using a Shapiro-Wilk test and homoscedasticity using a two-sample F-test.
Before the study, a sample size calculation had been performed as described by Chow et al. 42 for a one-sample design, given the mean (1.2°) and standard deviation (4.1°) of the CT-to-MR difference previously reported in the literature for CEA. 10 It resulted in a required sample size of 30 paired measurements for the CEA.   Table 1 reports the average values of mean error and surface distance obtained between the CT and sCT. The negative mean error indicated that the HU of bone on sCT was on average underestimated.

| Bone morphology
The average surface distance was below the image resolution with a submillimeter residual error as shown by the RMSE in Table 1. Errors were mostly located on the edge of the image, where less information is available, around the trochanter and around the ischium. A 360°view is available in Video S1.    Figure 4C shows the pairwise differences between the measurements performed by Reader 2 on CT and sCT on the 60 hip joints. For comparative purposes, measurements are displayed relatively to the intra-and interobserver variability. No patient presented considerable differences in all measurements which indicates that the overall morphology was conserved in sCT reconstructions.

| Hip joint morphometric parameters
The most important differences were observed for CEA, SA, AASA, and PASA, on patients with osteophytes ( Figure 4A,B).      adolescents and young adults. However, the purpose of this study was to assess the agreement between CT and sCT and we do not expect relevant differences in sCT generation between our study population and the target population. Sex-related changes in bone shape should not affect the model as it is a patch-based method, not prone to global morphological changes as demonstrated in a study performed in canines of various shapes and sizes that used a similar method. 29 Another factor that could potentially influence the voxelwise accuracy of sCT generation is bone density. However, as bone density is expected to be in the same range in elderly males and young adults, the intermodal differences should be similar between the two groups.

| Statistical analysis
In conclusion, sCT is a promising alternative to CT for the as-