Introduction

Positron emission tomography (PET) allows the interrogation of amyloid deposition in the brain, a hallmark of Alzheimer’s disease neuropathology¹, while magnetic resonance imaging (MRI) with its exquisite soft tissue contrast allows for imaging morphology-based features such as cortical atrophy, representative of neurodegeneration². These complementary strengths allow for MRI to assist in PET image processing and enhancement. While the absolute quantification of radiotracer concentrations is a strength in PET, radioactivity associated with the radiotracers will also present a risk to participants, especially in vulnerable populations. This will affect the scalability of large-scale clinical longitudinal PET studies.
With the advent of deep learning (DL)-based machine learning methods such as convolutional neural networks (CNNs), we have previously generated with DL diagnostic quality amyloid PET images using actual ultra-low injected radiotracer dose and simultaneously acquired MRI inputs³. Here, we will investigate the value of simultaneity in the acquisition of MRI inputs as well as the effect of different DL training methods in relation to these inputs. Non-simultaneous structural MRI inputs will allow for greater utility of MRI-assisted ultra-low-dose PET imaging to include those acquired on standalone PET/computed tomography (CT) and MRI machines.

Methods

PET/MR data acquisition
48 participants were recruited for the study. Participant recruitment was approved by the Stanford Institutional Review Board and all participants (or an authorized surrogate) provided written consent. 32 (19 female, 68.2±7.1 years) were used for pre-training the network; 328±32 MBq of the amyloid radiotracer [¹⁸F]-florbetaben were injected into the participants. 16 (6 female, 71.4±8.7 years) were scanned with the ultra-low-dose protocol. These participants were scanned in two PET/MRI sessions, with 6.49±3.76 and 300±14 MBq [¹⁸F]-florbetaben injections respectively. In both sessions the T1-, T2-, and T2 FLAIR-weighted MR images were acquired simultaneously with PET (90-110 minutes after injection) on an integrated PET/MR scanner (SIGNA PET/MR, GE healthcare) with time-of-flight capabilities (Figure 1). 7 participants were scanned in the same day (ultra-low-dose protocol followed by the full-dose protocol), while the others were scanned on separate days (1- to 42-day interval, mean 19.6 days). All images were co-registered using the software FSL to the standard-dose PET image.
CNN implementation
Using the low-dose CNN with multimodal PET/MRI inputs structure proposed in Chen et al. (Figure 2) ⁴, network training was either done from scratch (Method 1), or fine-tuned based on the participants presented in Chen et al.⁴. Either all layers of the U-Net were fine-tuned using the actual low-dose datasets (Method 2) or just the last layer (Method 3). 8-fold cross-validation was used to efficiently utilize all datasets (14 for training, 2 for testing per fold). For all methods, the training was carried out twice: once using PET/MRI inputs from the same scanning session (simultaneous: S) and once using PET and MRI from different sessions (non-simultaneous: NS) (Figure 1).
Data analysis
Using the software FreeSurfer, a brain mask derived from the T1 images of each subject was used for voxel-based analyses. For each axial slice, the image quality of the enhanced and low-dose PET images within the brain were compared to the full-dose image using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and root mean square error (RMSE). Paired t-tests were carried out at the p=0.05 level to test for significance of the metrics derived from NS vs. S inputs.

Results and Discussion

Qualitatively, all enhanced images showed marked improvement in noise reduction to the ultra-low-dose image and resembled the ground truth image (Figure 3). All three metrics showed significant image quality improvement (Figure 4) from the ultra-low-dose images to the enhanced images, and the metrics between the images enhanced from the S vs. NS inputs were not statistically significant with paired t-tests (Method 1: PSNR: p=0.27, SSIM: p=0.26, RMSE: p=0.15; Method 2: PSNR: p=0.99, SSIM: p=0.62, RMSE: p=0.83; Method 3: PSNR: p=1.00, SSIM: p=0.79, RMSE: p=0.67). The results showed that for this population and our interval between scanning sessions, simultaneity is not a strict requirement for morphology-based MR-assisted ultra-low-dose PET enhancement. The p-values from the t-test comparisons also showed that more similar metric values were obtained when more parameters were pre-trained and fixed (Method 3).

Conclusion

This work has shown that amyloid PET images can be generated using trained CNNs with both S and NS multimodal ultra-low-dose PET/MR images, broadening the utility of ultra-low-dose amyloid PET imaging.

Acknowledgements

This project was made possible by the NIH grant P41-EB015891, the Stanford Alzheimer’s Disease Research Center, GE Healthcare, the Foundation of the ASNR, and Life Molecular Imaging.