MRI-Based Attenuation Correction for Whole-Body PET/MRI: Quantitative Evaluation of Segmentation- and Atlas-Based Methods

Data Acquisition

Pairs of MR and PET/CT images from 11 patients were used. All patients had a clinical indication for an ¹⁸F-FDG whole-body PET/CT study for oncologic diagnosis. The patients were 6 men and 5 women (mean age ± SD, 66 ± 10 y).

¹⁸F-FDG PET/CT examinations were performed on a Biograph HiRez 16 (Siemens) with a low-dose CT scan using a slice thickness of 5 mm and a voxel size of 1.37 × 1.37 mm within the axial image planes. Details of the protocol can be found in the study by Brechtel et al. (17). Patients were asked to hold their breath in normal expiration during the acquisition of the CT data.

Complementary MR images were acquired on a 3-T Verio system (Siemens) using a dual-echo sequence (echo time 1, 1.23 ms; echo time 2, 2.46 ms; repetition time, 3.6 ms; flip angle, 10°). The isotropic spatial resolution of the images was 2.6 mm. In addition, a 2-point Dixon method (13) was used to generate fat- and water-only MR images. The reconstructed opposed phase images (echo time 1) were not used in this study. The acquisition time for each bed position was 18 s, which allowed for breath-hold acquisition.

All data were acquired with patients’ arms up in the PET/CT and MRI examinations. As shown in Supplemental Figure 1 (supplemental materials are available online only at http://jnm.snmjournals.org), patient positioning was similar in the PET/CT and MRI acquisitions. The study was approved by the local ethics committee.

Data Preparation

Marked intensity inhomogeneities are typical of whole-body MR images (Supplemental Fig. 1). Assuming that intensity inhomogeneity can be described by a multiplicative term (18), we corrected for this throughScorrected(x)=Smeasured(x)/b(x),Eq. 1where b(x) was a low-pass filtered version of the original image S_measured, obtained using a Gaussian kernel with a width of σ=5 cm. This approach amplifies the noise in the low-intensity region outside the patient. Therefore, we used a series of morphologic operations to segment the patient, thereby setting all values outside the patient to zero.

Accurately aligned MRI/transmission or MRI/CT image pairs are mandatory prerequisites for the evaluation of MRAC methods in general. Here, MR and CT images were aligned using a combination of a rigid registration algorithm based on the maximization of mutual information and a nonrigid registration method using a vector field obtained via the composition of small displacements, as previously described (19).

To improve the correspondence between MR and CT image volumes in the training atlas database, areas outside the field of view in the head and neck region of the atlas MR images and misregistered arms were automatically corrected for; they were replaced with segmented areas from the corresponding atlas CT image. Original uncorrected MR images were used to evaluate the MRAC methods.

To evaluate the importance of bone in general and assess the maximum accuracy that can be achieved using a no-bone approach to AC, we generated pseudo-CT images from the original CT images. Two sets of images were generated: in CT_nobone, all CT values greater than 200 Hounsfield units (HU) were set to 200 HU, leaving CT values of 200 HU or less unchanged. In CT_seg, the CT image was segmented into the 4 classes: air, lung, fat, and soft tissue.

Basic MR Image Segmentation (MRSEG) Approach

We performed a 5-class segmentation of the MR images into air, lungs, fat tissue, a fat–nonfat tissue mixture, and nonfat tissue using the in-phase, fat, and water images. For this purpose, we used image intensity–based thresholding and a simple analysis of connected areas of low intensity. Air and nonfat tissue were determined by thresholding the intensity-normalized in-phase MR image. The misclassification of low-intensity bone voxels as air was minimized via a morphologic image closing operation (10). On the basis of the Dixon fat and water images, voxels within fat images of more than double the intensity as compared with water images were assigned to fat and vice versa. For voxels with fat and water intensities within a factor of 2, we assumed a mixture of fat and water and assigned an attenuation coefficient identical to the mean of the nonfat and fat tissue attenuation. The lungs were detected as the largest connected group of low-intensity voxels.

This approach failed for 2 of 11 patients; in those cases, metal implants caused signal loss, and the low-intensity voxels in the lung appeared as if they were connected with the air outside the patient. We addressed this problem by manually adjusting the lung segmentation for these 2 patients. More advanced methods of lung segmentation in MR images should be capable of fully automating this step (20), but that was beyond the scope of our work. The locations of the metal implants can be found in Table 1.

For each of the 5 segmented classes, we assigned the average CT value derived from segmented CT images (Table 2), creating a pseudo-CT image PsCT_MRSEG.

AT&PR Approach

The AT&PR method is based on our previous work (6) but includes several modifications intended to assist in the processing of whole-body data. We used a different registration method and modified kernel function and added pre- and postprocessing steps. The nonrigidly registered MR/CT image pairs were used as an atlas and training database. In our evaluation, we used a leave-one-out cross-validation approach; that is, for each patient, we used the remaining 10 MRI/CT pairs as our atlas database. In the following, the phrase atlas database always refers to the complete database, excluding the images of the patient for whom we are predicting the pseudo-CT image.

In the first step in our process, MR images of the atlas database were registered to the patient MR image using Elastix (21). Images were registered using the 5-class segmentation. Low-intensity values in the abdominal region were assigned a value between those of fat and water to account for the variability in the number, shape, and position of gas pockets in the abdomen, which should be disregarded during the registration process.

The centers of gravity of the atlas and patient MR images were aligned to provide the registration algorithm with optimal initialization. The transformations of the atlas MR images to the patient MR image were applied to the corresponding atlas CT images. This process yielded 10 atlas-based pseudo-CT predictions for each patient.

In the second step, our method predicted a pseudo-CT value for each test voxel on the basis of nearby voxels in the atlas database, creating a pseudo-CT image PsCT_AT&PR.

Although averaging the registered atlas CT images would already yield a valid pseudo-CT image for a given test patient, anatomic variability between subjects is high, and nonrigid registration and subsequent averaging would smooth out patient-specific details. It is important to note that the AT&PR method not only uses the position of a voxel in the registered images but also matches the local structures of the MR images, facilitating more patient-specific prediction. Gaussian process regression (22) was used to predict a pseudo-CT value for each voxel of interest.

The kernel, which acts as a similarity measure, was chosen as follows:k(di,dj)=exp(−‖w(pMR,i)−w(pMR,j)‖22σMR,patch2)×exp(−‖w(pSeg,i)−w(pSeg,j)‖22σSeg,patch2)×exp(−‖xi−xj‖22σpos2).Eq. 2Here, d=(pMRT,pSegT,xT)T and pMR and pSeg are rectangular subvolumes (patches) from the in-phase MR image and from the segmented image, w is a weighting vector that assigns a higher importance to voxels in the center of the patch than to voxels further away from the center, and x is the center position of each training or test patch. σpos, σMR,patch, and σSeg,patch are parameters that determine how similarity in position, MRI patch, and segmentation image patch influence the overall kernel value. Optimal parameter values were determined using cross-validation.

The algorithm was trained on samples d that were drawn from random locations in the atlas database. A higher sampling density was used near bone regions because these are particularly challenging.

The pseudo-CT value for each voxel of interest was calculated usingcl=kTC−1y,Eq. 3where c_l denotes the predicted pseudo-CT value of a given voxel of interest l. k, with entries k_i = k(d_i,d_l), is a vector of kernel values from training patches d_i and test patch d_l, with an element of the matrix C_i,j = k(d_i,d_j) + σ²δ_i,j, and y is a vector of the CT values corresponding to the centers of the training patches d_i. σ² is a parameter for variance, which we empirically set to 50,000. A more detailed explanation of our method can be found in the study by Hofmann et al. (6).

Because the kernel decays quickly with the distance ‖xi−xj‖, we used only the nearest neighbors of the voxel of interest in the Gaussian process prediction formula. The nearest neighbor search was performed efficiently using TSTOOL (23).

In the final step, we applied the following automatic postprocessing rules, which override the prediction of AT&PR in cases in which the MRI intensity is sufficient to reliably determine the tissue class: if the local intensity in the in-phase image is higher than a given threshold and if the Dixon segmentation predicts fat (or water), then the voxel in the predicted pseudo-CT image is determined to have at least the value of fat (or water). Voxels that have low MRI intensity and for which none of the atlas-registered CT images predict bone in a nearby location are assumed to be air with a pseudo-CT value of −1,000 HU. This rule helps us to identify gas pockets even if, because of the high variability between patients, no similar structure is present in the atlas database.

Data Evaluation

PET images were reconstructed using the e7 tools (Siemens) via ordered-subsets expectation maximization, with 4 iterations and 8 subsets as in our standard clinical image processing with attenuation and scatter correction. PET image reconstructions were performed using either CTAC (for the reference cases) or MRAC. To create the MRI-derived attenuation maps, the attenuation coefficients of objects outside the patient (including the bed, positioning devices, and ancillary objects) were added to the attenuation map derived from the original CT images. Supplemental Figure 2 shows the original MR and CT images, resulting pseudo-CT images, and PET images obtained using the pseudo-CT data for AC for a representative patient.

PET images obtained via these methods of AC were compared using a set of spheric volumes of interest (VOIs). A nuclear medicine expert placed 21 VOIs on regions of normal physiologic uptake (norm regions) in whole-body PET/CT scans (Fig. 1). In addition, 50% isocontour VOIs were placed on areas of abnormally increased tracer accumulation relative to the surrounding tracer uptake (lesions).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

Mean relative errors for SUVs in PET images reconstructed using MRAC with reference images reconstructed using CTAC for VOIs in thorax (A), abdomen (B), and pelvis (C) and for lesion VOIs (D). Triangles indicate individual data points; circles and horizontal bars show means and SD.

To quantify the errors that would be observed in typical clinical settings, we evaluated the standardized uptake values (SUVs) in the PET images with CTAC and MRAC using both segmented and atlas-derived attenuation maps. The SUVs were calculated on the basis of body weight:ISUV(x)=I(x)×body weightInjected dose,Eq. 4where I(x) is the activity concentration at voxel x, and injected dose is the dose of ¹⁸F-FDG at the time of injection.

Of the 231 VOIs in standard anatomic regions, 223 VOIs were used for subsequent statistical evaluation, 7 were excluded, and 1 was evaluated separately because of artifacts caused by metallic implants. In addition, 28 VOIs were placed on lesions; of these, 22 were used for subsequent statistical evaluation, 4 were excluded, and 2 were evaluated separately. Of the 11 excluded VOIs, 3 were in abdominal gas pockets (which moved between MRI and CT acquisitions), 4 were in areas not properly aligned during MRI/CT registration, and 4 were in areas affected by severe breathing artifacts.