Unsupervised deep learning improves low dose SV2A PET imaging: a correlation study with motor severity in Parkinson’s disease

Introduction: Our recent dynamic ¹¹C-UCB-J PET study found reduced synaptic density in the substantia nigra (SN) and red nucleus (RN) regions in Parkinson’s disease (PD) patients. Dose reduction confers less radiation exposure in PET imaging, however, introduces more noise in the parametric images. Unsupervised deep image prior (DIP) methods reveal that convolutional neural networks (CNNs) can capture image features from corrupted images without training. The goal is to develop an adaptive DIP post-processing technique and study its effect on the correlation between synaptic density measured by reduced-dose dynamic PET and motor severity measured by the Movement Disorders Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) Part III score.

Methods: Twenty-one PD patients (9 males/12 females, age 62.8±8.9 yrs, weight 74.0±11.3 kg, UPDRS III 30.0±9.3) each underwent at least a 60-min dynamic ¹¹C-UCB-J PET scan on a HRRT (injected dose 592.4±144 MBq) to acquire the list-mode data. By redistributing the full-count data sequentially with 1 ms time step into 10 subsets, 10 replicates of 1/10-count datasets were generated. Both the full- and 1/10-count datasets were reconstructed into 21 frames including all degradation corrections using the MOLAR algorithm. To address the issue of overfitting to noise in DIP methods, we developed adaptive DIP denoising to automatically determine the stop iteration. Because low-frequency information is learned before noise during the DIP processing, a running variance with a window size of 100 iterations was calculated to detect the minimum variance, which corresponds to the selected stop iteration. A time-weighted composite image averaged over all frames was used as the network input to provide spatial information and the individual frame was used as the DIP target in our study. We performed kinetic analysis voxel by voxel using the simplified reference tissue model 2 (SRTM2) with the centrum semiovale as the reference region to estimate the binding potential (BP_ND) images from the full-count and 1/10-count frames without or with the adaptive DIP denoising. The ensemble normalized standard deviation (EnNSD) among 10 realizations of each patient and averaged over all patients was calculated as a measure of noise and the relative error (RE) with respect to the full-count estimate was used to measure the bias before and after denoising. For each PD patient, the BP_ND mean values were calculated in 11 primary regions of interest (ROIs) to measure the synaptic density. Correlations between the estimated synaptic density and motor examination were assessed using Pearson’s r at a significant level of 0.05.

Results: A significant negative correlation is found between the full-count BP_ND values and the UPDRS III scores in the SN (r=-0.52, p=0.02) and RN (r=-0.47, p=0.03 ). Lowering the dose to 1/10 resulted in a noise of 100% with a bias of 23% in the estimated BP_ND images and no significant correlation (r=-0.37, p=0.09 ) was found in any noise realization of the SN ROI. The adaptive DIP method reduced the noise to 38% with a bias of 22% while improving correlations of all realizations to significant (r=-0.47, p=0.03 ). With the noise of 74% and a bias of 15% in the RN of 1/10-count BP_ND images, there were 9 realizations showing no significant correlation (r=-0.39, p=0.08 ) between the regional mean and the UPDRS III scores. Reducing the noise (bias) to 28% (13%), the adaptive DIP method improves correlations of all replicates to significant (r=-0.46, p=0.04).

Conclusions: Reducing the injection dose of ¹¹C-UCB-J to 1/10 of the clinical level affected the correlation between the measured synaptic density and motor severity scores in PD related regions. We demonstrated that the proposed adaptive unsupervised DIP method was effective to enhance this correlation in the reduced-dose study. Statistical analysis with more subjects is ongoing and more conclusive results are expected.

• Download figure
• Open in new tab
• Download powerpoint