MR-Based Attenuation Correction Methods for Improved PET Quantification in Lesions Within Bone and Susceptibility Artifact Regions

Baseline Segmentation (SEG1)

The default method for attenuation map prediction on the Biograph mMR is based on segmenting the MR images into 5 classes with predefined constant LACs for each class, based on the method proposed by Martinez-Möller et al. (17). The class denominations and the corresponding LACs are outer air (0 cm⁻¹), lung (0.0224 cm⁻¹), fat tissue (0.0854 cm⁻¹), nonfat tissue (0.1 cm⁻¹), and a fat/nonfat tissue mix (0.0909 cm⁻¹). For this method, the vendor-provided software was used.

Refined Segmentation (SEG2)

An adapted segmentation method was proposed by our group (9) and further developed for this study. In the first step, a segmented MR image is created, which is used for subsequent registration and gaussian process regression (29) steps. The image is segmented into the classes outer air, lungs, fat tissue, nonfat tissue, fat/nonfat tissue mixture, and an additional class to denote regions of low MR signal intensity inside the body.

Fat and nonfat tissues as well as fat/nonfat tissue mixture were classified on the basis of the Dixon fat and water images. Voxels in the fat images of more than twice the intensity of water were assigned to fat and vice versa; remaining voxels that were not subsequently assigned to air or lungs are considered fat/nonfat tissue mixture. Outer air was defined as the largest connected region of low intensity that contained more than 50% of the volume edge voxels. For patients with metal implants—for example, sternal cerclages or portal catheters—the resulting artifacts can generate a connection of outer air and lung regions. To prevent the propagation of potential susceptibility artifacts inside the lungs, morphologic operations were applied to the inner part of the body contour to limit the propagation. Lungs were detected by hierarchic clustering as a set of 1 or 2 regions of low intensity, with the following constraints: first, the size of the regions had to be at least 15,000 voxels; second, the highest axial position in the body was required; and third, the difference in the centroid positions had to be less than 25%. The classes outer air, lungs, fat tissue, nonfat tissue, and fat/nonfat tissue mixture were assigned predefined attenuation values (0, 0.024, 0.0856, 0.1007, and 0.0988 cm⁻¹). The remaining class of low-MR-signal-intensity regions was introduced to account for the MR susceptibility artifact regions and consisted of potential artifact regions and nonlung, air-filled regions inside the body (e.g., sinuses, ear cavities, or air inside the gastrointestinal tract). This intermediate image is denoted MR_seg in Supplemental Figure 1. Thus, the subsequent atlas registration was less affected by the different positions of air inside the body or by the presence of susceptibility artifacts. For the creation of the final attenuation map, these regions were classified as soft tissue or inside air.

Atlas-Based Detection of Artifacts

Regions of low MR signal intensity inside the body were differentiated into air-filled regions that were not part of the lungs (e.g., air in the gastrointestinal tract) and soft-tissue regions masked by susceptibility artifacts. To accomplish this differentiation, external knowledge in the form of an atlas of potential artifact positions was used. Eleven patients (6 male, 5 female; age, 66 ± 10 y) were used for the atlas database. For each patient in the atlas database, a map of potential artifact locations was created. The transformations resulting from the registration of the atlas volumes to the patient volume were also applied to the artifact atlas maps. A map of potential artifact locations for the patient was then created by averaging the individual atlas artifact maps and performing a subsequent dilation operation. For every region of low MR signal intensity, the overlap with the artifact location map was computed. If the overlap was higher than a predefined threshold (90%), the low MR intensity of the region was assumed to have been caused by a susceptibility artifact. Then, the attenuation value of soft tissue was assigned to this region; otherwise, the attenuation value was set to air. Supplemental Figure 1 illustrates the workflow.

Atlas- and Pattern Recognition–Based Prediction (AT&PR)

The AT&PR method uses a database of previously aligned pairs of MR–CT volumes to predict attenuation values on a continuous scale (9,10). The prediction is based on gaussian process regression (29). For each voxel in the patient MR image, this method predicts a pseudo-CT value based on the spatially closest voxels from the images in the atlas database. Additionally, the local neighborhood of the MR and segmented MR images is matched, allowing for a more patient-specific prediction. The pseudo-CTs are converted to AC maps with a piecewise linear transformation (8). Hofmann et al. have presented a detailed description of the method (10) and an adaptation and evaluation for whole-body MRAC (9).

In the work presented here, we aimed to improve the robustness of the method in the presence of artifacts by the detection of MR artifacts on the basis of spatial location. In the first step, MR volumes were segmented as described in “Refined Segmentation (SEG2)” and “Atlas-Based Detection of Artifacts.” In the second step, the segmented MR volumes from the atlas database were nonrigidly registered to the patient volume using the open-source software Elastix (30). The resulting transformations were then applied to the corresponding CT images. In the third step, gaussian process regression was applied to predict a pseudo-CT value based on the spatially closest voxels from the images in the atlas database. By matching the local neighborhood of the MR and segmented MR images, a more patient-specific prediction was facilitated. The pseudo-CTs were converted to AC maps using a piecewise linear transformation. Finally, postprocessing steps were applied as described by Hofmann et al. (9).

Combining AT&PR and Segmentation-Based Prediction (SEG2wBONE)

For soft-tissue regions and lesions, differences in PET images reconstructed with either segmentation- or AT&PR-based attenuation maps are not considered to significantly affect PET image quantification (9,17,18,27). Consequently, the prediction of bone via the AT&PR method can be constrained to the parts of the attenuation map where osseous tissue can occur, and the impact of ignoring bone might become significant.

The areas of possible bone occurrence are determined from the coregistered atlas images. A location map of bone tissue for each coregistered atlas CT image is computed by applying an empirically determined bone/soft-tissue threshold (LAC > 0.1106 cm⁻¹), followed by gaussian filtering and a dilation operation. A combined bone location map is computed by averaging the individual bone location maps, similar to the computation of the artifact location map. Subsequently, the computation of the attenuation values via gaussian process regression is applied only to voxels inside the bone location map. Voxels with predicted LACs below the bone/soft-tissue threshold are assigned attenuation values from the MR segmentation described in “Refined Segmentation (SEG2)” and “Atlas-Based Detection of Artifacts.” Supplemental Figure 1 shows the workflow.

CT-Based Attenuation Correction (CTAC)

We used CTAC as the gold standard in our study. The CT-based attenuation maps were derived from CT volumes acquired during the PET/CT examination of the same patient.

The CT volumes were converted to attenuation maps and deformably registered to the segmented MR volumes created by the SEG2 method using Elastix. The initial alignment was based on the lung center of gravity. Beds, positioning devices, and ancillary objects were removed from the CT volumes before the registration.

Because the CT and MR volumes were acquired with different arm positions (in the arms-up position for PET/CT but with the arms down for PET/MR), the computation of the registration metric in the process of nonrigid registration was constrained to the body region and excluded regions with susceptibility artifacts. The arm regions in the CT-derived attenuation maps were replaced using the baseline SEG1 attenuation map created on the scanner. An example is shown in Supplemental Figure 2. The MR radiofrequency coils and the patient bed were automatically added during the reconstruction process.

Reconstruction

The PET images were reconstructed using an implementation of the 3D ordered-subsets expectation maximization algorithm on the Biograph mMR with 3 subsets and 21 iterations. Scatter correction was enabled. The image plane size was 172 × 172 voxels with a spatial resolution of 4.2 × 4.2 mm and a slice thickness of 2 mm. A 3D gaussian filter of 3 mm in full width at half maximum was applied.

Volume-of-Interest (VOI) Analysis

To quantify the differences between the attenuation maps, 3D 80% iso-contour VOIs were placed at the lesion locations on the PET images reconstructed with SEG1 attenuation correction. In 2 cases, lesions were located within an MR artifact. Here, the VOIs were placed on the PET images reconstructed with the AT&PR attenuation correction. VOIs smaller than 5 voxels were excluded from further evaluation. The VOIs were assigned to the categories soft tissue (n = 21), bone (n = 20), near bone (n = 19), and artifact (n = 9). These categories were evaluated separately. The distribution of lesions among the patients is shown in Supplemental Figure 3. Lesions were assigned to the category “near bone” if the maximal distance of the VOI border to the nearest bone tissue (LAC > 0.11 cm⁻¹) was below 30 mm but still outside the bone tissue in CT images. No head or neck lesions were included in the evaluation.

Relative differences between VOI mean values in the PET images after MRAC and reference CTAC were calculated and tested for statistical significance using a Wilcoxon signed rank test. Adjustment of the P values for false discovery was performed with the Benjamini and Hochberg method using the R environment for statistical computing (31). The results were considered statistically significant if the P value was less than 0.001.

For each method, a histogram of the attenuation values in the osseous VOIs was calculated for a sample patient and fitted with a gaussian distribution.