Zero-TE MRI-based attenuation correction for chest FDG PET/MRI with deep learning using unpaired PET/CT data: Quantitative analysis for difference in values of normal organ and malignant lesions

Purpose: The chest is challenging for deep learning (DL) approaches to generate pseudo computed tomography (CT) (pCT) data for attenuation correction (AC) of positron emission tomography (PET)-magnetic resonance imaging (MRI) due to differences in subjects’ respiratory phases, body positions, and scanner bed shape among training data. Unsupervised generative adversarial networks with adaptive layer-instance normalization for image-to-image translation (U-GAT-IT) combined with a modality independent neighborhood descriptor (MIND) yields pCT data generation without MRI data misregistration. We aimed to generate pCT data for AC in the chest with five-tissue classification including the bone component from zero echo time (ZTE) MRI by DL (MRAC-5) and assess differences with conventional MRI-based AC with four-tissue classification (MRAC-4) in CT Hounsfield units (HU) and standardized uptake value (SUV).

Methods: In total, 363 patients who underwent chest fluorodeoxyglucose PET/MRI with ZTE were retrospectively evaluated. ZTE images were scanned with large fields of view and reconstructed in-phase by central frequency adjustment. Unpaired training image data included bias-corrected ZTE (n=333) and CT components of PET/CT (n=333) and were utilized for training the U-GAT-IT/MIND model of pCT generation without manual annotations for training images. MRAC-5 was created by merging the segmented bone map from pCT onto a conventional 2-point Dixon-based AC map (MRAC-4) and was applied to PET reconstruction on the offline workstation. Thirty patients with malignant lesions who underwent ZTE PET/MRI and separate PET/CT on a different day were evaluated to quantitatively validate MRAC-5 in five regions in normal organs (lungs, fat, soft tissue, bone, and overall regions) and five regions with malignant lesions (mediastinum, lungs, chest wall, abdomen, and overall regions). The mean HU of pCT on MRAC-5, MRAC-4, and CTAC in normal organs and mean SUV corrected by MRAC-5 and MRAC-4 in normal organs and malignant lesions were individually assessed by measuring cubic regions of interest drawn in the respective regions. Bland-Altman plots were used to compare differences in the mean HU between MRAC-5 and CTAC and between MRAC-4 and CTAC. To evaluate the differences in mean SUVs between MRAC-5 and MRAC-4, Wilcoxon’s signed rank test was performed to assess differences in the normal organs in five regions and in the malignant lesions in five regions.

Results: The mean differences in HU in overall regions between MRAC-5 and CTAC [-31.00; 95% confidence interval (CI), -39.34 to 22.65] were significantly smaller than those between MRAC-4 and CTAC (-70.64; 95% CI, -88.82 to -52.46) (p<0.0001). The difference was statistically significant in bone (p<0.0001) but not in the fat, liver, and lungs (p>0.05). The SUVs in the normal organ were significantly higher in MRAC-5 than those in MRAC-4 in soft tissue, bone, and overall regions (p<0.0001, <0.0001, and =0.0006, respectively). The SUVs in malignant lesions were significantly higher in MRAC-5 than those in MRAC-4 in mediastinal, abdominal, and overall regions (p=0.0005, 0.0039, and 0.0006, respectively). Conclusion: ZTE-based AC maps with bone classification generated by DL yield smaller differences of HU on pCT with CTAC and larger SUVs in normal organs and malignant lesions than those with conventional MRAC, suggesting the feasibility of DL-based training using unpaired datasets to generate AC maps of chest PET/MRI.

• Download figure
• Open in new tab
• Download powerpoint