Introduction

A major difficulty in the application of body PET/MRI is accurate attenuation correction, which is of importance for PET quantification¹. An area of interest for PET/MRI attenuation estimation is the lung region, for which current methods usually assign a uniform attenuation coefficient. With the better characterisation of lung tissue in ZTE images, we hypothesise that ZTE images can be used to generate an attenuation map for the lung and improve the performance of current MR attenuation correction algorithms in this area.

Methods

Six participants underwent thoracic CT examinations (120 kVp; 10-30 mAs; pitch: 1.375; voxel size: 0.98×0.98×3.24 mm³) and ⁶⁸Ga-DOTATATE^_PET/MRI (SIGNA PET/MR, GE Healthcare, Waukesha, WI) on two separate occasions. The manufacturer's ZTE (radial) pulse sequence was used to obtain 3D images of the chest: TE: 16 µs, TR: 228 ms, flip angle (FA): 1^o, voxel size 1.9×1.9×2.6 mm³, RBW: ± 244kHz, NEX: 4, field-of-view (FOV): 50 cm. In and out-of-phase images were acquired using a dual echo 3D Dixon sequence: TE₁: 1.11 ms, TE₂ : 1.67 ms, TR: 4.05 ms, FA: 5_­^o, voxel size 1.9 x 1.9 x 2.6 mm³, RBW: ±166 kHz, NEX: 0.7, FOV: 50 cm. Water and fat separated images were subsequently calculated.
For each participant, the ZTE images were co-registered with the Dixon images using rigid transformation. Before bias correction and intensity normalisation, principal components analysis was applied to the Dixon images to extract the first two principal components. Lung segmentations were subsequently performed, using a five-cluster k-means algorithm with the ZTE images and the two Dixon principal components as inputs. Bias correction and normalization utilized a hybrid approach², which applied a surface-fitting³ and a histogram-based method⁴ to the body and lung regions, respectively.
CT images were co-registered to the ZTE-Dixon image pairs using deformable registration. Conversion of ZTE images to a pseudo-CT (pCT) for the lung regions was achieved by fitting a multiple linear regression model with the bias-corrected/normalized ZTE image and two Dixon principal components as inputs, and the CT image as output. The performance of the model was validated using a leave-one-out strategy.
Following pCT generation, CT and pCT images were converted to attenuation maps using bilinear transformation⁵. Two hybrid attenuation maps were created by substituting the lung region of the Dixon attenuation map with values from the CT and pCT to generate Dixon/CT and Dixon/pCT attenuation maps, respectively.
PET images from 50-80 min post injection of ⁶⁸Ga-DOTATATE were reconstructed using time-of-flight ordered-subsets expectation-maximization (TOF-OSEM): 256×256×89 matrix; 2.34×2.34×2.78-mm voxels; 2 iterations/28 subsets. Data corrections were applied as implemented on the scanner, with a 4-mm FWHM Gaussian filter post reconstruction. For each participant, three different reconstructions were produced using the Dixon/CT, Dixon/pCT, and manufacturer’s (Dixon) map for attenuation correction. Regions (cylinder with 5 mm² circular area and 2.78 mm height) were placed in lung tissue on the differently reconstructed PET images of each participant in the anterior, posterior, superior, inferior directions. Differences between the attenuation correction techniques were quantified as the % change in the standardized uptake value (SUV) in these regions, using the Dixon/CT as the reference for comparison (values report as median [interquartile range]). The Wilcoxon signed rank test was used to statistically compare data with non-normal distribution.

Results

Figure 1 shows the joint histogram of normalized ZTE vs CT intensities in all participants. Figure 2 demonstrates CT, normalised ZTE, pCT images, and absolute difference images between the CT and pCT in the lung. Figure 3 illustrates representative orthogonal views of a scanner-generated attenuation map (Dixon; 3A-3C), together with the corresponding hybrid attenuation maps, Dixon/CT (3D-3F) and Dixon/pCT (3G-3I). Figure 4 shows images reconstructed using different attenuation maps. Figure 5 presents boxplots of regional SUV and % differences to the Dixon/CT method.
Compared to Dixon/CT, the Dixon/pCT and Dixon-based attenuation correction methods resulted in median errors of 1.2 [0, 6.46]% and -4.26 [-10.0, 0]% respectively. The Dixon method resulted in an underestimation of SUV in anterior (-9.72 [-13.0, -6.98]%) and posterior regions (-7.38 [-10.5, -4.17]%) of the lung as compared to Dixon/CT. In contrast, use of the Dixon/pCT overestimated SUV in these areas (anterior vs posterior: 7.69 [4.35, 8.33]% vs 2.08 [0, 4.76]%). Both Dixon and Dixon/pCT produced similar errors for regions located inferiorly in the lung (Dixon vs Dixon/pCT: 3.1 [0, 4.17]% vs 0.6 [-4.55,6.9]%), while SUV were overestimated by the Dixon method in regions in the superior direction.

Discussion

Attenuation correction using the Dixon/pCT tended to overestimate SUV in the lung as compared to Dixon/CT. An underestimation of SUV was observed with the Dixon-based method. Overall, the Dixon/pCT method resulted in an absolute error of 3% in the lung when compared to the Dixon-based attenuation map. Errors between methods in regions placed in the superior or inferior portions of the lung were comparable. However, a negative bias in the anterior-posterior direction was observed, consistent with the gravitational dependence of lung density in this direction for a participant lying supine on the scanner bed.

Conclusion

A hybrid ZTE/Dixon method for MR-based attenuation correction was developed and applied to the lung. Combination of ZTE-based and Dixon-derived attenuation maps reduced the SUV errors in the lung by 3%, showing potential for use in PET/MR attenuation correction.

Acknowledgements

This study is supported by the Wellcome Trust (211100/Z/18/Z).