UTE‐mDixon‐based thorax synthetic CT generation

Abstract

Purpose

Accurate photon attenuation assessment from MR data remains an unmet challenge in the thorax due to tissue heterogeneity and the difficulty of MR lung imaging. As thoracic tissues encompass the whole physiologic range of photon absorption, large errors can occur when using, for example, a uniform, water‐equivalent or a soft‐tissue‐only approximation. The purpose of this study was to introduce a method for voxel‐wise thoracic synthetic CT (sCT) generation from MR data attenuation correction (AC) for PET/MR or for MR‐only radiation treatment planning (RTP).

Methods

Acquisition: A radial stack‐of‐stars combining ultra‐short‐echo time (UTE) and modified Dixon (mDixon) sequence was optimized for thoracic imaging. The UTE‐mDixon pulse sequence collects MR signals at three TE times denoted as UTE, Echo1, and Echo2. Three‐point mDixon processing was used to reconstruct water and fat images. Bias field correction was applied in order to avoid artifacts caused by inhomogeneity of the MR magnetic field. Analysis: Water fraction and R2* maps were estimated using the UTE‐mDixon data to produce a total of seven MR features, that is UTE, Echo1, Echo2, Dixon water, Dixon fat, Water fraction, and R2*. A feature selection process was performed to determine the optimal feature combination for the proposed automatic, 6‐tissue classification for sCT generation. Fuzzy c‐means was used for the automatic classification which was followed by voxel‐wise attenuation coefficient assignment as a weighted sum of those of the component tissues. Performance evaluation: MR data collected using the proposed pulse sequence were compared to those using a traditional two‐point Dixon approach. Image quality measures, including image resolution and uniformity, were evaluated using an MR ACR phantom. Data collected from 25 normal volunteers were used to evaluate the accuracy of the proposed method compared to the template‐based approach. Notably, the template approach is applicable here, that is normal volunteers, but may not be robust enough for patients with pathologies.

Results

The free breathing UTE‐mDixon pulse sequence yielded images with quality comparable to those using the traditional breath holding mDixon sequence. Furthermore, by capturing the signal before T2* decay, the UTE‐mDixon image provided lung and bone information which the mDixon image did not. The combination of Dixon water, Dixon fat, and the Water fraction was the most robust for tissue clustering and supported the classification of six tissues, that is, air, lung, fat, soft tissue, low‐density bone, and dense bone, used to generate the sCT. The thoracic sCT had a mean absolute difference from the template‐based (reference) CT of less than 50 HU and which was better agreement with the reference CT than the results produced using the traditional Dixon‐based data.

Conclusion

MR thoracic acquisition and analyses have been established to automatically provide six distinguishable tissue types to generate sCT for MR‐based AC of PET/MR and for MR‐only RTP.

1. Introduction

Accurate attenuation correction (AC) for PET/MRI whole‐body scans and MR‐only radiation treatment planning (RTP) for the thorax remains an unresolved challenge as most the approaches focus on body regions other than the thorax. Specifically, multi‐atlas deformable image registration and deep learning‐based methods were proposed for brain,^{1, 2} head and neck,^{3, 4, 5} and pelvis^{6, 7, 8, 9, 10} sCT generation. Compared to these regions, the thorax is more complicated and heterogeneous. The complicated thoracic structures limit the accuracy of deformable registration for the multi‐atlas approach and reduce the reliability of reference CT for deep learning‐based training. For these reasons, although showing good performance for brain, head and neck, and pelvis sCT generation, those approaches may not be a good solution for thoracic sCT generation. In fact, thorax is a problematic anatomical area given respiratory and cardiac motion artifacts¹¹ and because it contains numerous tissue types including bone, air, fat, water/blood, and muscle that span a wide range of CT HU values and MR signal intensities. Furthermore, bone has fast T2* relaxation and air/lung has low‐proton density which causes both to have a similarly low MR signal and thus poor differentiation when using conventional MRI pulse sequences.¹² Technically, a precise PET AC with an error below 5% requires the identification of at least six tissue types, that is, cortical bone, spongeous bone, soft tissue, lung, air, and fat.^{13, 14, 15} Nevertheless, identification of the requisite tissue types is still elusive, especially because chest tissues span the entire physiologic range of photon attenuation.

With this perspective, we intended to identify six tissue types, that is, air, lung, fat, soft tissue, spongeous (low‐density) bone, and dense bone. We saw the problem as one of creating a synthetic CT (sCT) that can be used instead of a measured CT for PET/MR image reconstruction or MR‐based RTP.^{16, 17}

The ultra‐short echo time (UTE) MRI sequence was proposed as a potential approach for thoracic imaging.¹⁸ With a combination of an echo time (TE) less than 1 ms and a fast radiofrequency (RF) pulse excitation, the UTE technique simultaneously accounts for the short T2, low‐proton density, and motion artifact of the lung¹⁹ and it supports the discrimination between air and bone.²⁰ In this study, UTE was combined with modified Dixon (mDixon) reconstruction, that is, UTE‐mDixon, to provide additional anatomic localization and differentiation between fat and water. Furthermore, the multi‐echo, UTE‐mDixon sequence data were spatially encoded using radial stack‐of‐stars (SoS) sampling in order to facilitate the anterior‐to‐posterior dimension to be adjusted to a patient’s chest size, and thereby achieving time‐efficient coverage of the thorax. The acquisition angular sampling density and reconstruction voxel size were optimized using phantom and normal volunteer data. To the best of our knowledge, no UTE‐mDixon acquisition has been proposed and optimized for the thoracic scans. Furthermore, a feature selection process was performed to determine the most informative SoS UTE‐mDixon image combination for automatic thoracic tissue clustering so that the proposed thoracic MR sequence will be able to access the information provided by both UTE and mDixon techniques with a single, free‐breathing acquisition to generate thoracic sCT for MR‐based PET AC and for MR‐only RTP.

2. Materials and methods

2.1. Data acquisition and reconstruction

We used a free breathing coronal, multi‐echo, stack‐of‐stars UTE‐mDixon sequence for thoracic imaging (Fig. 1). The UTE‐mDixon sequence includes a free‐induction decay signal acquired at an ultrashort TE time (0.14 ms) as well as two echoes (TE = 1.14 and 2.14 ms). A modified three‐point Dixon reconstruction, which performs water and fat separation with three MR signals not exactly collected at water/fat in‐phase and opposed‐phase TEs, was used.²¹

Figure 1.

A pulse sequence plot of the stack‐of‐stars ultra‐short‐echo time‐mDixon sequence.

As the quality of the final UTE‐mDixon image strongly depends on the angular sampling density of the radial acquisition, it was optimized for the proposed method. The angular sampling was varied based on the percentage of sampling density with 100% defined as the total number of k‐space samples of a fully‐sampled Cartesian encoding. Due to the fact that the radial sampling has dense sampling in the center of the k‐space and coarse sampling in the high‐frequency region, compared to a Cartesian encoding, the radial sampling needs to be 157% of the Cartesian sampling points, that is, π/2 times of the Cartesian sampling lines, in order to fulfill the Nyquist sampling requirements.²² It can also be estimated that a radial sampling scan requires two times the sampling points to yield the same total number of samples per‐unit‐area of a Cartesian scan. For these reasons, the optimal angular sampling was searched from 150% to 200% to cover the sampling range resulting in high‐quality images. In addition to the proposed UTE‐mDixon acquisition, a traditional diagnostic, breath hold, dual‐echo mDixon scan was performed as a standard reference for comparison. The detailed scan parameters are listed in Table 1.

Table 1.

MR Acquisition parameters of the mDixon and the ultra‐short‐echo time‐mDixon data

	Scan 1	Scan 2	Scan 3	Scan 4
Sequence	mDixon	UTE‐mDixon	UTE‐Dixon	UTE‐mDixon
Scan Percentage	100%	150%	175%	200%
Scan Direction	Axial	Coronal	Coronal	Coronal
Trajectory	Cartesian	Radial	Radial	Radial
Field of View RL/AP/FH (mm)	458/458/302	458/302/458	458/302/458	458/302/458
TR (ms)	2.8	4.6	4.6	4.6
TE (ms)	0.86/ 2.66	0.14/1.14/2.14	0.14/1.14/2.14	0.14/1.14/2.14
Slice thickness (mm)	2.9	2.9	2.9	2.9
Flip angle in degrees (°)	5	10	10	10
Number of slices	104	104	104	104
SENSE acceleration	2.0	1.5	1.5	1.5
Total scan duration (s)	16.7	213.3	248	284
Breath Control	Breath hold	Free breathing	Free breathing	Free breathing

An MR ACR phantom was scanned using both MR sequences for comparative assessment of the image quality. Spatial resolution was measured. Image uniformity was measured using the percent integral uniformity (PIU) based on the guideline of the American College of Radiology.²³ PIU is calculated using the intensities of both a high region and a low region in an ACR slice with a uniform material as follows:

$$<![CDATA[\text{PIU}\left(\%\right)=\left[\mathrm{1}-\left(\frac{high-low}{high+low}\right)\right]\times \text{100},]]>$$ (1)

In addition to the phantom scan, 25 healthy volunteers were recruited (13 females and 12 males). The age of the population was 27.5 ± 12.1 yr; body weight was 68.4 ± 14.6 kg; and chest circumference was 92.9 ± 9.2 cm. All of the volunteers provided written informed consent for the IRB‐approved protocol. All MR examinations were acquired using the MR of a Philips Ingenuity TF PET/MRI (Philips Healthcare, Amsterdam, Netherlands) which is essentially an Achieva 3.0 T MRI. Each participant was scanned in a supine, arms‐down position, and underwent four thoracic scans, as listed in Table 1, using a SENSE 16‐channel torso TX multi‐transmit XL surface coil with CLEAR (Constant LEvel AppeaRance).

2.2. Data processing and feature estimation

The MR data were reconstructed into three primary images, that is, UTE, Echo₁, and Echo₂. Furthermore, modified, three‐point Dixon reconstruction was used for water and fat separation yielding Dixon_water and Dixon_fat images. Three different voxel sizes, that is, 2.9 × 2.9 × 2.9 mm³, 1.9 × 2.9 × 1.9 mm³, and 1.45 × 2.9 × 1.45 mm³, in the left–right × anterior–posterior × inferior–superior directions, were used in order to evaluate the sampling effect of the reconstruction. The mDixon data were reconstructed using the voxel size that is suggested by the manufacturer, that is, 0.96 × 0.96 × 2.9 mm³.

In addition, owing to the B₀ and B₁ field inhomogeneity introduced by a person in an MR scanner, the reconstructed images have a position‐dependent tissue intensity. This artifact degrades the MR image quality, results in the inconsistency of the signal for a given tissue type, and degrades the accuracy of the tissue classification.^{24, 25} For this reason, a multiplicative, intrinsic component optimization was used for the bias field correction for UTE, Echo_1, and Echo₂ images,^{26, 27} and the results were compared to those without the uniformity correction.

The corrected UTE, Echo₁, Echo₂, Dixon_water, and Dixon_fat images were used to calculate Water_fraction, calculated as Dixon_water/(Dixon_water + Dixon_fat). Furthermore, R2*, that is, 1/T2*, can be used to provide bone information and was estimated from the data collected at different TEs using the Eq. (2):

$$<![CDATA[I(k)=I_{0}exp\left(-R{2}^{\ast }\times TE(k)\right),]]>$$ (2)

in which I(k) is the MR signal intensity at the kth TE series and I₀ is the initial MR signal strength after an RF excitation.¹² R2* was estimated using the non‐linear least‐squares fitting with the non‐negativity constraints in order to avoid negative R2* values caused by the noise in the lung region.

2.3. Feature selection for data clustering

In order to maximize the performance of fuzzy c‐means (FCM) clustering used for automatic tissue classification,^{12, 28} a feature selection process was used to determine the most informative, non‐redundant MR feature combination. Seven candidate MR features were considered, that is, UTE, Echo₁, Echo₂, Dixon_water, Dixon_fat, Water_fraction, and R2*. As a preprocessing step, all feature values were normalized to a range between 0 and 1.^{29, 30} The feature selection process proceeds to find the best three‐feature combination as there are originally three, independent MR signals in the data, that is, UTE, Echo₁, and Echo₂. There were 35 (seven choose three) possible three‐feature combinations drawn from a set of seven possible features. In order to select the optimal feature combination, we calculated, for each three features and resultant classes, the fuzzy hypervolume (FHV). FHV is defined as:

$$<![CDATA[FHV=\sum _{j=1}^C{\left(\frac{{\sum }_{i=1}^N{\mu }_{\mathit{ij}}^m\left({\mathbf{x}}_i-{\mathbf{c}}_j\right){\left({\mathbf{x}}_i-{\mathbf{c}}_j\right)}^T}{{\sum }_{i=1}^N{\mu }_{\mathit{ij}}^m}\right)}^{1/2},]]>$$ (3)

where x _i is the feature vector for voxel ith, N is the total number of voxels, c _j is the centroid of the jth class, C is the total number of classes, $<![CDATA[{\mu }_{\mathit{ij}}^m]]>$ is the membership (probability) functions of the voxel i and class j, and m is the fuzziness parameter which was set to 2 in this study. To cover a range of different clusters, the total number of classes C varied from 4 to 8. As such, each subject would generate 175 (35 combinations × 5 numbers of clusters) FHV values. For each feature set, the average FHV among the 5 numbers of clusters and among all the subjects was used as a clustering performance measure. Since a smaller size of all classes in the feature space corresponds to more robust clustering results,^{31, 32} the best feature combination would therefore result in the smallest FHV.

2.4. Automatic tissue classification and sCT generation

The steps of the automatic tissue classification are summarized as:

• Step 1: Determine air outside and inside the body separately;

• Step 2: FCM clustering;

• Step 3: Determine initial tissue types using the map in Fig. 2;

• Step 4: Determine the lung mask and resolve dense bone and lung tissues using the lung mask;

• Step 5: sCT generation

Figure 2.

The proposed automatic tissue assignment map. The values of Dixon_water and Dixon_fat were used to determine the final tissue type of each class. [Color figure can be viewed at wileyonlinelibrary.com]

In Step 1, the air voxels outside a subject's body are identified as those having values less than a 15% threshold of maximum in an image that is the sum of the UTE, Echo₁, and Echo₂ images. As air inside a subject's body, for example, air in the stomach, may be affected by the surrounding tissue, a different approach is needed. Specifically, internal air voxels are identified as those with Water_fraction values lower than 10% of their maximum and which are located inside the subject's body.

In Step 2, voxels inside a subject's body, which are determined in Step 1, are used for clustering. FCM clustering group voxels with similar signal properties into six classes. This number was selected as yielding the best bone separation in preliminary studies.

In Step 3, the initial tissue type of each class is determined according to the FCM centroid location of each class. Cluster centroids appearing in the areas marked as “Fat,” “Low‐density Bone,” or “Soft Tissue” in Fig. 2 are simply assigned as these types. A centroid located in the “Low‐density Bone (or fat)” region is assigned as “Low‐density Bone” if there is no other cluster that has its centroid in the “Low‐density Bone” region; otherwise the cluster is assigned as fat. Centroids located in the “Dense bone + Lung” have to be subsequently separated, voxel‐wise, as detailed in the next paragraph.

Step 4 uses spatial information to resolve dense bone and lung. As lung voxels can be determined from their high R2* values due to susceptibility effects, an initial lung mask is created from R2* values that exceed 80% of the maximum. Holes in this mask are filled using morphologic, flood‐fill operation (MATLAB Image Processing Toolbox function imfill). A voxel association with the “Dense bone + Lung” cluster is interpreted as being lung if the voxel lies within this lung mask and otherwise as dense bone. After all the tissue classes are determined, automatic morphology processing was applied on the FCM membership functions to reduce clustering errors introduced by the respiratory motion and the MR susceptibility effect in order to improve the accuracy of the sCT. At the completion of Step 4, each voxel has values indicating memberships with each cluster which we interpret as being fractional composition of fat, soft tissue, low‐density bone, and either dense bone or lung.

In Step 5, voxel values in the sCT are computed as a weighted sum of the HU values of the component tissue types.¹² The HU values of air, body air, fat, water, soft tissue, low‐density bone, and dense bone are set to −1000, −850, −98, 0, 40, 385, and 657, respectively.³³ Special HU value assignment for the component tissue types was found to be required for voxels within the lung mask. Lung was set to −850 HU. As the remainder of the voxel, mainly blood and blood vessels, were assumed to be best represented as water, the HU value of 0 was assigned for other classes when calculating the weighted sum. The final sCT field of view (FOV) is confined to a circle within 90% of the total coronal FOV, that is, a circle with a diameter of 412.2 mm (458 mm × 90%) in the coronal view, in order to avoid artifacts close to the edge of the FOV.

2.5. Reference CT (rCT) generation

While a measured CT scan would be the ideal reference for validating the sCT generation method, it was not appropriate to expose the normal volunteers to the associated ionizing radiation. Accordingly, we used a template‐ or model‐based method using normal, average human structure, to create a reference CT (rCT) for testing the proposed method. The template method has been shown to be able to generate accurate thoracic CT for PET AC and radiation treatment plans for people with normal anatomy^{34, 35} and would be applicable for normal volunteers in order to create a reference for comparison. Male and female XCAT computational human models were used in this study for the template CT generation.³⁶ Using XCAT, a library of XCAT CT and Dixon_water MR was generated. To account for the mathematical morphology of the bone skeleton in XCAT CT (which did not segment bone marrow³⁷), bone from de‐identified measured CT data replaced the bone skeleton in the XCAT so that the final reference CT would contain accurate bone anatomy.

In short, rCT generation is done as follows. First, template MR is registered to the volunteer's MR including manual, landmark‐based 3D b‐spline registration³⁸ and diffeomorphic local‐phase registration.³⁹ Second, the CT bone model of the template CT would be transformed to match the volunteer's MR according to the deformation field estimated by the MR‐to‐MR registration. Finally, the whole XCAT template CT would be warped to match MR in order to generate the final rCT for evaluation. In this study, the sCT was designed for MR‐based PET AC. For this reason, a 4‐mm isotropic Gaussian filter was applied on both sCT and rCT for resolution matching.^{40, 41}

2.6. Performance evaluation

In order to evaluate the accuracy of the proposed method, the mean error (ME; sCT ‐ rCT), mean absolute error (MAE), and correlation coefficients (R) of voxels in the body mask, that is, voxels inside the body contour, and those in the bone mask obtained by taking a threshold of 100 HU on the rCT were calculated. This threshold was defined using a separate cohort of patients who had CT scans but no MRI in order to include low‐density bone structures in low‐dose CT scans (data not shown). The difference between the results of different methods was evaluated using the Student’s paired t‐test with a significance level of 0.05 for hypothesis tests.

3. Results

3.1. Image quality assessment

The MR ACR phantom was scanned using both the traditional mDixon pulse sequence and the proposed coronal SoS UTE‐mDixon sequence. Figures 3(a) and 3(b) shows that bias field correction improves the consistency of the MR signal across the field of view. After correction, the pixel values in the center region are not higher than those in the outer region. The measured PIUs of the UTE‐mDixon and the mDixon images are shown in Fig. 3(e). The mDixon image has higher uniformity than the UTE‐mDixon image. The bias field correction improves the uniformities of the images acquired by both sequences and makes the uniformity of the UTE‐mDixon image close to that of the mDixon image. The ripple artifacts are reduced when the sampling density is getting higher (Fig. 4) and are negligible when the sampling density is 200%.

Figure 3.

Image quality assessment of the mDixon and the ultra‐short‐echo time (UTE)‐mDixon scans using the MR ACR phantom. The UTE‐mDixon images without and with the bias field correction are shown in (a) and (b), respectively. In order to further evaluate the spatial fidelity of the proposed UTE‐mDixon, the mDixon image with bias field correction is shown in (c) and is further fused with the UTE‐mDionx image (b) using checkerboard display (d). Furthermore, the percent integral uniformities (PIU) of the mDixon and the UTE‐mDixon images with different sampling densities and reconstructed voxel sizes are shown in (e). The term “BC” indicates the results with the bias field correction.

Figure 4.

The sagittal ultra‐short‐echo time images of different sampling densities, that is, 150% (a), 175% (b), and 200% (c). The ripple respiratory artifacts can be seen in 150% and 175%, but not in 200% images.

Also regarding uniformity, the UTE‐mDixon image with 200% sampling density yielded the highest uniformity of all sampling densities, and the acquisition can be done in 4.7 min. The uniformity decreases when the reconstruction voxel size is reduced, but the effect is not as prominent as that of the sampling density.

Regarding the spatial resolution of the UTE‐mDixon images, Fig. 5 shows the Echo₂ images reconstructed using different voxel sizes. The images using the 1.45 × 2.90 × 1.45 mm³ voxel size are sharper than those at other voxel sizes. For this reason, the spatial resolution was measured using the sharpest, 1.45 × 2.90 × 1.45 mm³ images. The spatial resolutions of mDixon and UTE‐mDixon images were measured by analysing the differentiated step response function of the ACR phantom. The measured in‐plane, full‐width‐at‐half‐maximums of the differentiated step response functions for the mDixon and the UTE‐mDixon images are 2.81 and 3.27 mm, respectively.

Figure 5.

The Echo2 of the ultra‐short‐echo time‐mDixon data reconstructed using different voxel sizes. The data were reconstructed using 2.90 × 2.90 × 2.90 (a), 1.90 × 2.90 × 1.90 (b), and 1.45 × 2.90 × 1.45 mm³ (c); the (d), (e), and (f) show the corresponding zoomed images of the circled regions. [Color figure can be viewed at wileyonlinelibrary.com]

These results suggest that the image with 200% sampling density, bias field correction, and 1.45 × 2.90 × 1.45 mm³ voxel size would be the best choice for the proposed thoracic SoS UTE‐mDixon imaging. Therefore, these scanning parameters were used for the sCT generation.

3.2. Feature selection

Figure 6(a) shows the results of the FCM‐based feature selection. The FHV was used as the indicator of the robustness of tissue clustering and was used to determine the best feature combination. As shown in the three‐feature results of Fig. 6, the best feature combination is the one with Dixon_water, Dixon_fat, and Water_fraction, as it achieved the lowest FHV. Images of the selected features are shown in Figs. 6(b)–6(d). Moreover, Figs. 6(e) and 6(f) show the zoomed lung Water_fraction images of the traditional mDixon and that of the UTE‐mDixon data and illustrate that the UTE‐mDixon images reveal lung tissue details that are not visible in mDixon images.

Figure 6.

The results of the feature selection and the selected features. The feature selection using the fuzzy hypervolume (FHV) is shown in (a). The features are labeled using different numbers, that is, 1: ultra‐short‐echo time (UTE); 2: Echo1; 3: Echo2; 4: Dixon_water; 5: Dixon_fat; 6: Water_fraction; and 7: R2*. The smaller the FHV, the better the feature combination. The dashed box indicates the best feature combination, that is, Dixon_water (b), Dixon_fat (c), and Water_fraction (d). In order to compare the lung images acquired by different sequences, a zoomed lung Water_fraction images of traditional mDixon and that of UTE‐mDixon data are shown in (e) and (f). The sizes of the lung are not the same in (e) and (f) as the mDixon and the UTE‐mDixon images were acquired using breath‐hold and free‐breathing techniques, respectively. The image (g) shows the zoomed lung R2* map of the UTE‐mDixon data. [Color figure can be viewed at wileyonlinelibrary.com]

3.3. Automatic tissue clustering and classification

Figure 7 shows the results of the FCM clustering. Panel (a) shows the initial tissue centroids. The tissue type of each centroid was determined based on the map in Fig. 2. Figure 7(b) shows the final adjusted tissue centroids in the selected 3‐feature domain after the lung and body masks were applied (Step 4). With one exception, all of the tissues of the 25 volunteers were reviewed and were correctly identified using the proposed automatic tissue classification scheme. As shown in Fig. 7(b), the centroids of different tissue types are well‐separated, except for a soft‐tissue centroid and a low‐density bone tissue centroid, that is, the triangle close to the low‐density bone centroids and the cross symbol overlapped with the fat centroids. The miss‐located tissue centroids were from an obese (BMI > 30 kg/m²) volunteer who has chest and waist circumferences of 115 and 124 cm, respectively.

Figure 7.

The tissue clustering results of the proposed method. The initial tissue centroids of all of the subjects based in Fig. 2 are shown in (a). The final adjusted tissue centroids determined by the automatic tissue assignment process are shown in (b). The final data clusters are separated into different tissue types. The estimated tissue spatial distributions of the six tissue types, that is, air (c), lung (d), fat (e), soft tissue (f), low‐density bone (g), and dense bone (h), are shown. (c)–(g) show the same coronal slice of a subject. (h) shows the slice where the dense bone of the spine is visible. [Color figure can be viewed at wileyonlinelibrary.com]

The estimated six tissue distributions of air, lung, fat, soft tissue, low‐density bone, and dense bone tissues are shown in Figures 7(c)–7(h). Except for (h), the same coronal slice of a subject is shown in (c)–(g); (h) shows the slice where the dense bone of the vertebrae can be seen. The estimated spatial distributions can indicate the location of different tissue types, for example, the edge and shape of lung. Therefore, the tissue spatial information can be further used for tissue localization.

3.4. sCT generation and evaluation

Figure 8 shows a direct comparison between sCT and rCT. Both the results of the voxels in the body mask and those in the bone mask are shown. The UTE‐mDixon‐based sCT results in a MAE <50 HU for the body mask. The MAE is lower if all of the voxels in the whole volume, including air in the background, were used for the calculation, that is, 24.1 ± 5.2 HU, although we chose not to score it this way as getting the vast majority of the background voxels correct is easy and dilutes the effect of the improvement. In addition, the error maps of the corresponding volunteer are shown in Fig. 8(b). The mDixon‐based sCT shows a large error in the lower lung due to the difference between the lung shapes of the breath‐hold and free‐breathing techniques. For the bone tissue, the bone voxels of the mDixon‐based sCT are underestimated as the bone voxels would be treated as soft tissue. Overall, the UTE‐mDixon‐based sCT shows smaller errors than the mDixon‐based sCT. However, some errors can be seen in the lung region of the UTE‐mDixon‐based sCT, while the mDixon‐based sCT has few errors in the corresponding regions.

Figure 8.

A comparison between sCT and rCT. (a) shows the mean errors (ME), mean absolute errors (MAE), and correlation coefficients (R) of the body and bone masks of all of the volunteers. The error bar indicates the standard deviations. The differences between mDixon and ultra‐short‐echo time (UTE)‐mDixon‐based synthetic computed tomographies (sCTs) are significant in all of the groups (P < 0.0001). The images of sCT, reference CT (rCT), and the error map are shown in (b). The upper row of the sCT and sCT‐rCT maps corresponds to the results of the mDixon‐based sCT; the lower row is the results of the UTE‐mDixon‐based sCT. [Color figure can be viewed at wileyonlinelibrary.com]

4. Discussion

MR‐based sCT generation for PET AC and MR‐only radiation therapy planning has been a challenging problem, especially for the thorax and for patients having pathologic anatomy. To the best of our knowledge, we are the first to report on a MR‐signal‐based method for sCT generation in the thorax and we have done so in a way that does not use a multi‐atlas approach and therefore can properly accommodate abnormal tissue structures. The same problem can also be seen in deep‐learning based sCT generation methods. The training‐based deep learning methods might exhibit erratic behavior for predicting sCT when data are not present in the training data due to the complexity of thorax imaging. To directly account for thoracic complexity, we developed an MR‐signal‐based UTE‐mDixon technique that can be efficiently acquired for sCT generation. The proposed acquisition accounts for the rapid T2* decay of lung by acquiring data at three TE times, including the ultra‐short TE time (0.14 ms). Compared to the well‐established mDixon, our method better captures the lung and bone morphology visualization [Figs. 6(e)–6(g)], while preserving the spatial resolution and uniformity.

The proposed SoS UTE‐mDixon sequence uses a radial readout and may be subject to greater geometric distortion and off‐resonance artifacts arising from gradient‐induced eddy currents, as compared to the mDixon sequence which uses a Cartesian readout. As shown in Fig. 3, our initial investigation of the two sequences showed that the off‐resonance artifacts were decreased for the UTE‐mDixon sequence when the bias correction was applied. Furthermore, both sequences had minimal geometric distortion, Fig. 3(c). The individual as well as the checkerboard fused views of the two sequences acquired with the standard grid structure in the ACR phantom demonstrate that there are no distortions in the grid pattern within an individual acquisition or between the two sequences. These initial results indicate that the sequences did not introduce additional geometric distortion. However, given the additional sources of potential geometric inaccuracies, specifically, patient‐induced B0 distortions and gradient non‐linearity, an investigation regarding the overall geometric accuracy should be performed in order to ensure the robustness of using the proposed method for MR‐based radiation therapy.

While seven MR features are available for each voxel, our clustering approach was based solely on three features, that is, Dixon_water, Dixon_fat, and Water_fraction. This selected feature combination had the lowest FHV among other combinations and which translated in a robust estimation and a clear separation among clusters in the feature space.^{31, 32} In addition to relevance, our feature selection methods also account for redundancy. For example, UTE information seems to be embedded in the three‐point Dixon results [Fig. 6(f)] and can be extracted in the clustering process even if the UTE feature is not included in the final, selected features.

Subsequently, our two‐phase clustering approach yielded an accurate sCT by generating well‐separated tissue clusters [Fig. 7(b)]. The fact that the mislocated soft tissue and low‐density bone centroids occurred in one large volunteer suggests the possible limitations of our methods due to patient size. However, imaging in an obese population is an omnipresent problem among all technologies and as our algorithm worked without issue in the other large volunteers, we expect that the method can be made robust given sufficient data.

The proposed UTE‐mDixon‐based sCT shows a MAE less than 50 HU in the thorax. Clinically, these results would translate in accurate quantification of MR‐based PET AC and MR‐only RTP. Although the template‐based approach has been shown to be able to generate sCT with good quality,^{34, 35} having only the template‐based CT may have resulted in under‐ or over‐estimation of the errors. For example, the higher error in the lung region of the UTE‐mDixon‐based sCT compared to the mDixon‐based sCT [Fig. 8(b)] may not truly be an error of the proposed method but rather an artifact of our assessment caused by the lack of lung details in the rCT. In addition, the template‐based, also known as a model or atlas, method assumes that a patient's body structure is similar to that of the normal population. This assumption may be inaccurate when applying it to a patient with diseased tissue and body structure, for example, patients with pathologies of pulmonary edema, tumors or excised tissues. When applied to patients with pathologies, the template‐based method may result in incorrect CT HU values in the sCTs. In this study, we proposed an MR‐signal‐based technique. It does not require that the size, shape, and location of tissues be similar to those of the normal, healthy volunteers. The CT HU value of each voxel is determined based on the MR voxel signal. A tissue with a pathology with altered tissue composition would have correspondingly altered MR signal and consequently altered HU values in the sCT. Therefore, the proposed method would be applicable for patients with pathologies.

It is worth noting that the MAE was calculated using the template‐based CT as the reference. As such, the results may not be able to be directly compared to the MAEs of methods using measured CTs for validation. Our future work will focus on comparing accuracy of the sCT‐based PET AC and RTP dosimetry using measured CT data in patients receiving clinical PET/CT scans and who therefore would not receive extra ionizing radiation for research purposes.

5. Conclusions

We proposed a complete sCT generation scheme, including the new MR acquisition method, automatic tissue classification, and sCT formation, for thoracic imaging. The images generated by the proposed UTE‐mDixon method provide a wide FOV and good image quality compared to those of the traditional mDixon method. Furthermore, the proposed method generates spatial distributions of different tissue types which can be used for tissue localization and sCT generation. Consequently, the proposed method would be able to facilitate various kinds of clinical applications, including tissue localization, MR‐bead PET AC, and MR‐only RTP.

Conflicts of Interest

The work was funded by an NIH academic‐industrial partnership grant to Case Western Reserve University with Philips Healthcare as the industrial partner (R01 CA196687, PI Muzic) and with University Hospitals Cleveland Medical Center and Jiangnan University as clinical and academic partners. Co‐authors have associations with these entities as indicated in their affiliations that are listed on the title page.