Patient-specific transmission volume synthesis for attenuation correction in simultaneous PET/MR

Objectives: Attenuation correction in simultaneous PET/MR continues to present a challenge given the lack of a straightforward relationship between photon attenuation and MRI signal. This is especially problematic in psychiatric studies which require high fidelity PET quantification. Several solutions have been proposed, including the generation of patient-specific CT data from a preexisting atlas, or the use of specialized MRI pulse sequences to allow the 3-way segmentation of soft tissue, air and bone. Whereas some feel these strategies are sufficient for clinical use¹, current techniques are non-ideal for quantitative research in several ways: CT attenuation maps require interpolation in order to compensate for energy differences between CT and PET, while MRI techniques require additional scan time and generally assign assumed attenuation coefficients to different classes of tissue and therefore do not account for inter- and intra-patient variations in bone density². We utilized a patch-based³ algorithm to generate patient-specific transmission data, the historical gold standard for attenuation correction, from a preexisting atlas.

Methods: The algorithm utilizes an atlas consisting of 8 pairs of co-registered MPRAGE and transmission volumes acquired on separate scanners. The similarity between an input MPRAGE volume and each MPRAGE volume in the atlas is determined using patch-based weighting, obviating the need for deformation and incorporating an extent of nonlocal similarity. The output is a linear combination of patches in the transmission domain weighted by the similarity of the MPRAGE volumes. One patient’s data were removed from the atlas and used for validation. PET reconstruction was performed using filtered back projection.

Results: The output transmission volume showed a high degree of similarity when compared to the scanner data (Figure 1). In particular, the mean percent error between the scanner and synthesized data was -0.9%. Filtered back projection was used to reconstruct the PET data using both attenuation maps (Figure 2), the mean percent error between the reconstructions was 2.1%. The associated voxel-wise standard deviation was 13.3% (Figure 3). Discussion: The chief limitation of this procedure is the quality of transmission and MRI data used in the atlas, as they were collected in an older study without consideration of this purpose. The field of view of the MRI inconsistently crops the anterior- and superior-most areas of the head. An ongoing study using standalone PET and MRI at our institution will allow the construction of a higher quality atlas. Moreover, the original atlas could only make use of MPRAGE data, while the future atlas will allow us to investigate the usefulness of a range of pulse sequences. There is relatively high standard deviation of errors between the reconstructed PET images, although we believe this is likely due to the high noise of scanner-acquired transmission data relative to the ostensibly smooth synthesized data, as well as the algorithm’s automatic assignment of zero attenuation to any voxels outside of the head. The chief limitation of this method is shared among all atlas-based techniques, namely that it is generally limited in use to adult patients with relatively standard anatomy of the head.

Conclusion: Synthesizing transmission data from T1-weighted MRI can be a useful technique for simultaneous PET/MR studies which require accurate radiotracer quantification in the brain. Further study in our improving atlas, along with examination of sources of standard deviation in the results, could allow for improvement over current research techniques for PET/MR attenuation in the head. Research Support:

• Download figure
• Open in new tab
• Download powerpoint