Accurate Magnetic Resonance (MR) imaging based attenuation correction is crucial for quantitative Positron Emission Tomography (PET) in simultaneous MR/PET imaging. However, due to a lack of robust MR bone imaging methods, MR based attenuation correction remains a critical issue in MR/PET image reconstruction. In this work, we developed and evaluated a deep learning (DL) based MR based attenuation correction method for improved MR/PET quantification accuracy in prostatic cancer imaging.
Materials and Methods:
CT and PET/MR images were acquired in a cohort of 17 prostate cancer patients (mean age 65.2 ± 8.5 years; range 46 - 76 years). All patients underwent PET/CT (GE Discovery 710) and PET/MR (Siemens 3T Biograph mMR) studies (approved by the ethics committees of Monash University), using the protocols below.
PET/CT Acquisition: CT datasets of the patients were acquired with image resolution of 1.367 x 1.367 x 3.2 mm3, with an operating energy range of 120 – 140 kV. For the PET scan the average administered dose of Ga-68 was 183 ± 27 MBq (range 140 – 251 MBq).
PET/MR Acquisition: The PET/MR data acquisition were performed approximately 60 mins after the PET/CT scans, including the following sequences: Dixon with resolution 2.604 x 2.604 x 3.12 mm3, TR = 3.96 msec, TE = 1.23 msec, flip angle 9°. Thirty minutes of PET list-mode data were acquired and images were reconstructed using Siemens e7tools software, with ordered-subsets expectation maximization (OSEM) algorithm, 3 iterations, 21 groups, and post- smoothed using a 3mm Gaussian filter.
Data preprocessing: CT images were registered to the DIXON images using the ANTS tool. Attenuation correction μ-maps were calculated by converting CT numbers of the tissues to linear attenuation coefficients4.
Deep learning training: A variation of U-net5 was designed and trained to estimate four tissue classes (air, fat, soft tissue, and bone) in the μ-maps. The input to the network were the Dixon images including water, fat, in-phase and out-phase images. The reference output for the network was the set of co-registered CT μ-maps. Thirteen datasets were used to train the network, and four datasets were used for evaluation. The DL based MR based attenuation correction workflow is presented in Figure 1.
Generation and comparison of μ-maps: The attenuation coefficients (μ-values) for bones, soft tissue, fat and air were set to 0.151 cm-1, 0.1 cm-1, 0.092 cm-1, and 0.0 cm-1, respectively. The DL based method was compared with the following μ-maps:
[(2) and (3) were generated from Siemens MR/PET scanner]
Dice coefficients across all subjects for air, fat, soft tissue and bones were calculated for each of μ-map method. The PET image reconstructed using μ-mapct was taken as the reference, denoted as PETref. The difference between each μ-map and the reference was calculated for comparison.
Results:
Compared with the reference μ-mapct (see Figure 2A) the μ-mapdl demonstrated more accurate classification of the bones compared with μ-mapdixon+bone where the overall bone structures were significantly underestimated, as confirmed by visual inspection and comparison of DICE coefficients (see Table 1). Across the four testing patients, in the bone region the averaged DICE coefficient was 0.491 for μ-mapdixon+bone, and 0.818 for μ-mapdl, respectively. To evaluate the impact of the attenuation map differences the PET images were reconstructed using the three μ-maps (see Figure 3) and the corresponding error images were calculated (see Figure 4). When compared with the reference PET image reconstructed using μ-mapct, the PET image from μ-mapdl produced the most accurate PET quantification in the prostate region, resulting in only 1.6% error, whereas the image reconstructed using μ-mapdixon+bone had 8.4% error and the image from μ-mapdixon demonstrated a larger error of 13.4%.Discussion and conclusion:
In this work, we have presented an MR based attenuation correction method for PET using a DL approach. The DL method shows significant improvement in identification of bones for PET attenuation correction. Significantly improved PET quantification accuracy was demonstrated using MR/PET data in a prostate cancer patient study.