Comparison of Template-Based Versus CT-Based Attenuation Correction for Hybrid MR/PET Scanners

Abstract

Attenuation correction (AC) of cerebral PET data acquired in hybrid MR/PET scanners is still a challenge. To overcome this problem we previously proposed a correction method by obtaining template-based attenuation maps (AM) using MR and ECAT EXACT HR+ transmission scans. In the present study we investigated (a) the basic difference between template-based and CT-based AC methods and (b) their influence on reconstructed PET images. The data of 11 subjects undergoing $^{18}{\rm FDG}$ imaging in the Siemens 3T MR-BrainPET scanner were used. Additionally, from all participants a CT scan of the whole head was acquired at the same day. These CT images were transformed to CT-based AMs. They were filtered by a 3D Gaussian kernel with 3 mm (BrainPET resolution) filter width, which was considered as reference. Comparisons between both AMs (CT-based and template-based) were performed by estimating the Dice coefficients D and calculating the numbers of true positive, true negative and false negative voxels. The BrainPET emission data were reconstructed with both AMs ( ${\rm AM}_{{\rm CT}3{\rm mm}}$ and ${\rm AM}_{\rm Template}$ ). All reconstructed PET images were scaled to standardized uptake values (SUVs) and normalized to the MNI brain for using the AAL-VOI Atlas analysis. Correlation plots with regression equation, coefficients of determination ${{\rm R}^2}$ and relative differences (RD) between ${\rm AM}_{{\rm CT}3{\rm mm}}$ and the ${\rm AM}_{\rm Template}$ were derived. The fraction of the overall true positive voxels averaged over the 11 subjects was $80.4 \pm 7.5\% $ for ${\rm AM}_{\rm Template}$ compared to ${\rm AM}_{{\rm CT}3{\rm mm}}$ ( ${{\rm D}_{\rm bone}} = 0.63 \pm 0.08$ ; ${{\rm D}_{\rm soft - tissue}} = 0.85 \pm 0.08$ ; ${{\rm D}_{\rm air}} = 0.79 \pm 0.04$ ). A misclassification of bone as soft tissue and vice versa was evident in the comparison. The correlation plot of all VOIs considered (1,276 values) showed a mean ${{\rm R}^2}$ of 0.964 and a slope of 1.02. A mean RD of $1.33 \pm 0.95\%$ ( ${\min} = - 0.12\% $ , ${\max} = 2.85\% $ ) was found. The template-based AC method proposed by our group shows considerable differences in comparison to the higher resolution CT-based AM with respect to the Dice coefficients, in particular in the classification of bone and soft tissue. However, this has no major influence on the reconstructed $^{18}{\rm FDG}$ PET data.