Data-driven Head Motion Correction in PET Imaging Using Deep Learning

点击翻译按钮获取中文摘要

Project Summary/Abstract In the parent R21, we are developing deep learning (DL)-based head motion estimation models, based on the PET raw data, to track head motion during a PET scan in real time without the need for external motion sensors. In this supplement, we will pursue the development of deep learning neural networks dedicated to estimating motion for Alzheimer's disease (AD) subjects. Brain PET imaging is highly sensitive to head motion. Problems due to head motion are exacerbated by the long duration of the scans, with scan times commonly over one hour, and by the small scale of disease-focused regions of interest, e.g., hippocampus, for AD subjects. The Yale PET Center recently acquired a set of AD PET data that includes AD patients under treatment using CT1812, a first-in-class drug that displaces Aβ oligomers bound to neuronal receptors at synapses. In the CT1812 study, AD patients underwent baseline and post-treatment scans using 11C-UCB-J and 18F-FDG. The longitudinal nature of this study requires the detection of small-scale changes in small-scale AD-related brain areas over time within the same individual. Existing Polaris Vicra motion tracking has a 5-10% failure rate, therefore, there is a compelling need to develop accurate head motion correction for this study. In this administrative supplement, we will pursue the development of DL neural networks dedicated to estimating motion for the AD PET dataset acquired under the CT1812 study, and perform rigorous evaluations. In Aim 1, we will develop a novel DL methodology to perform motion correction, which includes: (1) a DL model to generate synthetic AD PET images based on rapid back-projection images for every 1-sec frame, and (2) a second DL model to estimate the rigid motion between two synthetic AD PET images. We will evaluate our motion estimation models using the data from the twenty subjects acquired in the CT1812 study against Polaris Vicra motion tracking. In Aim 2, we will perform kinetic modeling analysis for all the CT1812 studies for both tracers. Dynamic motion corrected reconstruction will be performed using the DL estimated motion correction (from Aim 1) and be compared to reconstruction using Vicra-based motion correction. We will correlate the changes in synaptic density (11C-UCB-J), glucose metabolism (18F-FDG) and cognitive function following CT1812 treatment. We hypothesize that our proposed DL-based approach will outperform the Vicra- based approach by reducing cross-subject variations within cohorts for any quantitative PET measure in both 11C-UCB-J and 18F-FDG tracers. We also hypothesize the DL-based approach will outperform Vicra by increasing absolute correlation coefficient value for any correlation between changes in PET measures and cognitive improvement.