Attenuation correction of Total Body PET using syntetic CT derived from the emission data.

Introduction: With the emergence of total body PET/CT such as the Biograph Vision Quadra, Siemens Healthineers, and the uEXPLORER, United Imaging, PET sensitivity has increased markedly allowing for reduction of the injected tracer dose by a factor of ten or more (Alberts et al, EJNMMI 2021). This has potential implications for easing the path to scanning subjects where a reduced radiation dose is desired; pediatric patients, pregnant woman, healthy controls etc. Moving towards PET scanning with very low patient radiation dose, the associated attenuation correction CT (AC CT) scan will become the primary source of ionizing radiation. We propose a deep learning driven synthetic CT (sCT) generation procedure where the sCT is produced directly from the PET non-attenuation corrected (NAC) images thereby eliminating the need for a separate AC CT scan on the long axial field of view (LAFOV) Biograph Vision Quadra.

Methods: Training: Paired NAC and CT images from 702 retrospective patients scanned on a short axis field of view (SAFOV) Siemens Biograph Vision 600 PET/CT with (n=602) and without (n=82) IV contrast were used to train the model based on Generative Adversarial Networks (GAN). As IV contrast introduces a known bias in the AC map, and thereby in the PET signal, the patient data were utilized in two separate steps in order to train a model generating sCT images without this bias while taking advantage of the full dataset. Firstly the generator, a standard U-Net, was trained using the 602 patients with IV contrast. Next, both the generator and discriminator were trained using the 82 patients without IV contrast. In both steps, a train/validation split of 80/20 % was applied to tune the hyper parameters and prevent overfitting of the networks.

Test: The independent test data set consisted of 13 patients scanned on the LAFOV scanner using a diagnostic PET/CT protocol. All patients received 3MBq/kg 18F-FDG 60 min prior to PET scanning. PET data was reconstructed using a clinical protocol of 5i4s, 2 mm Gauss filter, PSF and TOF modeling.

Evaluation: Attenuation corrected PET images were reconstructed using both the clinical AC CT (120 kVp, Ref mAs 160 mAs) and the sCT images. We report mean and maximum PET uptake in selected regions, lung, liver, pelvic muscle near bone, aorta. For a qualitative image quality assessment, all 13 PET test dataset were evaluated blinded by an experienced nuclear medicine specialist indicating preference for or indifference between the CT or sCT reconstructed images.

Results: Quantitative analysis of PET data revealed mean relative SUV deviations of lung, liver, pelvic muscle, aorta of across all selected regions of 2.4 ± 7.7%, 2.5 ± 3.7%, 4.7 ± 3.7%, 2.2 ± 4.4%, resp (range:[-13%,+18%]). The clinical evaluation showed no clinical difference in all 13 cases. In two cases a slight chance in contrast was noted. No preference between the sCT or CT based AC was found. No artifact was found in the PET data. Figure shows a representative patient: A: FDG PET attenuation corrected using CT. B: FDG PET attenuation corrected using deep learning derived sCT.

Conclusions: We present a GAN-based network that allows for reconstructing PET data from NAC data without the need for a CT. The method was directly applicable to data from the LAFOV scanner, despite being trained on data with 5-10 times less sensitivity. The method allows for reduction of a LAFOV PET scan such as the Quadra to less than 1 mSv in radiation dose by combining the high PET sensitivity with the sCT image for AC without loss of image quality. We hypothesize that certain typical errors in the PET images derived from wrong attenuation values could be addressed by this NAC-derived approach.This includes streaking artifacts, bias due to contrast agents, external motion between CT and PET or internal motion such as breathing leading to the typical shadows at the diaphragm region. Furthermore NAC based AC can have applications where CT is not available such as PET/MR or PET-only systems.

• Download figure
• Open in new tab
• Download powerpoint