Commercial PET/MR systems have recently been implemented in clinical environments, and serve promising results for the assessment of brain disease. One inherent drawback of PET/MR systems is the difficulty to obtain accurate attenuation correction (AC). One of the clinical PET/MR scanners, the GE SIGNA PET/MR, implements a single atlas-based method (s-Atlas). This method is comparatively accurate in supratentorial regions, while not accurate enough in the infratentorial regions. One possible solution is the multi atlas-based method (m-Atlas). Using multiple atlas datasets is expected to improve the accuracy of attenuation correction because it largely compensates the error from registration and patient variability. The aim of our study was to evaluate the feasibility of implementing a multi atlas-based method by using head FDG-TOF-PET/MR data, and by comparing it with the clinical s-Atlas, and with the gold standard method (CT attenuation correction method (CT-AC)).

We enrolled 12 patients. The median patient age was 62 years [range 31 to 80]. All patients underwent a clinically indicated whole-body 18F-FDG-PET/CT (GE Healthcare Discovery 690 PET/CT) for staging, re-staging or follow-up of malignant disease. All patients volunteered for an additional PET/MR scan of the head (GE Healthcare SIGNA PET/MR) (no additional tracer being injected). For each patient, 3 AC maps were generated. Both s-Atlas and m-Atlas AC maps were generated from the same patient-specific LAVA-FLEX T1-weighted (T1w) images, being acquired by default on the PET/MR scanner during the first 18s of the PET scan. The scan parameter is as below; TR ~ 4 msec, TE 2.23 msec, flip angle 5°, slice thickness 5.2 mm with 2.6mm overlap, 120 slices, pixel size 1.95 × 1.95 mm². S-Atlas AC map was extracted by the PET/MR scanner, and m-Atlas AC map was created using a web service　（http://goo.gl/0xIUYs , translational imaging group, University College London） which automatically generates m-Atlas pseudo-CT images. For comparison, the CT-AC map generated by PET/CT was registered and used as gold standard. Using each AC map, PET images were reconstructed from raw data on the TOF-PET/MRI scanner. All PET images were normalized to the SPM5 PET template, and FDG accumulation was quantified in 67 volumes-of-interest (VOIs; automated anatomical labeling (AAL), atlas). In each VOI, FDG uptake values from CT-AC (PET_CT) and each s-/m-Atlas (PET_{s-/m- Atlas}) were measured. Relative (%diff) and absolute differences (|%diff|) between images based on each atlas AC and CT-AC were calculated. For assessing the error distribution in the brain, the 67 AAL-VOIs were merged into 7 more generalized VOIs: frontal lobes, occipital lobes, parietal lobes, insula and cingulate gyrus, central structures (caudate nucleus, putamen, pallidum and thalamus), temporal lobes and cerebellum. FDG uptake in all VOIs and generalized merged VOIs were compared using paired t-test and Bland-Altman test.All 12 patients successfully underwent PET/CT and PET/MR examinations. Linear regression showed PET_CT and each PET_s-/m-Atlas　to be highly correlated (R2 > 0.99) to a straight line with a slope of 0.99. The Bland-Altman plot for all 804 VOIs proved that m-Atlas has no bias and no underestimation or overestimation (0.09 ± 1.52%; range -4.98% - 4.09%). These results are superior to the result of s-Atlas (0.16 ± 1.89%; range -5.00 – 4.84%). The superiority becomes clearer on the scatter plot. The average |%diff| of 804 VOIs with m-Atlas was significantly smaller than s-Atlas by 30% (s-Atlas vs. m-Atlas; 1.5±1.1% vs. 1.2 ± 0.9%, p < 0.01). The box plot of each generalized VOI shows that the underestimation with s-Atlas was pronounced in regions close to the skull base, such as temporal lobes and cerebellum. Notably, the %diff with m-Atlas in these regions was significantly smaller (s-Atlas vs. m-Atlas; temporal lobes, 1.31±1.38% vs. -0.38±1.46%, p < 0.01; cerebellum, 1.46±2.12% vs. -1.07±1.87%, p < 0.01). In the scatter plot of each generalized VOI, different degrees of positive correlation were found between s-Atlas and m-Atlas results with a slope of 0.307 – 0.721. Representative cases are given in figure 5.Errors introduced by using multi atlas-based attenuation correction method on a TOF-PET/MR scanner did not exceed 5% in any brain region. The error of PET images based on multi-atlas method is reduced from 1.5% to 1.2% compared to the single-atlas method, a 30% improvement. The greatest improvement with multi atlas-based attenuation correction was found in brain regions close to the skull base.