Abstract
108
Objectives: Using [11C]UCB-J PET imaging to measure synaptic vesicle protein 2A (SV2A), our recent study provided the first evidence of synaptic loss in brainstem nuclei involved in the pathogenesis of Parkinson disease (PD) in living patients. Lowering tracer dose reduces radiation exposure to patients however increases imaging noise, which could affect the separation of PDs from healthy controls (HCs). We developed a patch-based artificial neural network (ANN) and demonstrated its effectiveness in reducing noise while introducing minimal bias in PET images reconstructed from reduced count data. The goal of this study is to enhance the separability between PD and HC groups in reduced-dose SV2A imaging using a cascade ANN based denoising technique.
Methods: Six pairs of demographically matched PD and HC subjects underwent PET scans on an HRRT. The list mode data were acquired for at least 60 min after the start of [11C]UCB-J administration. The full-count dataset of each subject was down-sampled by redistributing the coincidence events sequentially with 1 ms time steps, resulting in 10 noise realizations of 1/10-count datasets. Every set of the full-count or 1/10-count data were reconstructed into 21 dynamic image frames with degradation corrections using the MOLAR algorithm. We developed an ANN model and trained it with the 1/10-count and full-count dynamic image frames of another randomly selected HC subject. The ANN model served as a nonlinear multivariate regression function to form a mapping between the training inputs consisting of 100k 3D patches extracted from a 1/10-count frame and the desired outputs of the corresponding patches extracted from the full-count frame. Mimicking anisotropic diffusion filtering, the trained ANN model was applied multiple times in cascade to each of the 1/10-count dynamic image frames of all the PD and HC subjects, resulting in the ANN processed 1/10-count dynamic image frames. We used the simplified reference tissue model 2 (SRTM2) with the centrum semiovale as the reference region to estimate the binding potential (BPND) map from the dynamic frames. For PD assessment, the BPND mean values were estimated in 11 primary regions of interest and the BPND reduction between the HC and PD groups were then computed for each ROI. We conducted two-tailed t-tests for 6 HCs versus 6 PD patients of the full-count and of each 1/10-count noise realizations with and without the ANN processing to evaluate the difference between the HC and the PD groups. Results: From the full-count data, significant group (HC versus PD) differences were found of the BPND values in the substantia nigra (SN) (41% reduction from HC to PD, p < 0.001) and red nucleus (RN) (-23%, p = 0.01). The significant group differences remained in the SN (-36%, p < 0.05) of all 10 noise realizations after the current dose was reduced to 1/10. With an average of 20% BPND value reduction in the RN, there were however 7 noise realizations of the 1/10-count data showing no significant differences between the HC and PD groups. After the ANN processing, the number of noise realizations without significant differences reduced to 2 with the p values being 0.17 and 0.33, respectively. These two noise realizations had relatively small regional BPND value reduction and large subject variation compared with other noise realizations. Conclusions: Reducing the tracer injection dose to 1/10 of the clinical level affected the separation of HC and PD subjects based on the BPND values in the RN. The cascade ANN method significantly enhanced the separation between these two groups in the reduced dose imaging. Statistical analysis on more subjects with applying the cascade ANN method to other levels of reduced dose is ongoing. More conclusive results will be reached in terms of the [11C]UCB-J PET imaging dose reduction for PD patient assessment.