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 Positron-emission tomography (PET) is an imaging modality that allows clinicians and researchers to study the physiological or pathological processes of the human body, and in particular the brain via the use of specific tracers. For brain PET imaging, patient head movement during scanning presents a challenge for accurate PET image reconstruction and subsequent quantitative analysis. Problems due to head motion are exacerbated by the long duration of the scans, with scan times commonly over one hour. Furthermore, some PET studies specifically involve subjects that either have trouble staying still due to psychological variations, e.g. patients with neurodegenerative disorders such as Alzheimer's disease and Parkinson's disease, or psychological variations, e.g. subjects with anxiety disorders, or are required to participate in tasks that involve movement, e.g. smoking cigarettes while scanning. In brain scans, the average head motion can vary from 7 mm in clinical scans to triple this amount for longer research scans. Quantitatively, a 5 mm head motion can produce biases of up to ~35% in regional intensities and ∼15% in volume of distribution estimates, which could much larger than the difference observed in regional intensities or binding potential that distinguish different demographic groups being studied. The ability to track and correct head motion, therefore, would be of high utility in both clinical and research PET studies. In the past, many motion correction methods have been proposed. However, except for hardware-based approaches, there has been no method that can track frequent head motion on-the-fly during the PET acquisition. Hardware-based approaches are not readily available for clinical translation or used by other research facilities due to highly-customized software/hardware setup. To address this challenge, we propose to develop a data-driven methodology using deep learning to track and estimate rigid head motion using PET raw data, and incorporate both tracer type and time as conditional variables into this deep neural network design in order to handle diverse PET tracer types and their dynamic behavior. Overall, these solutions will provide for a data-driven motion estimation methodology to improve the quality of PET imaging. Specifically, we will start with the development and testing of our methodology for rigid head motion estimation using single-tracer PET raw data. Then we will perform evaluation of our multi-tracer motion estimation methodology applied to real PET data with a diverse range of tracers. Finally, in the exploratory phase, we will integrate time-of-flight information into deep learning-based motion prediction. The significance of this proposal is that it will allow for improved quality of PET imaging in real time and potentially allow for its use in clinical PET systems that do not have special motion tracking hardware. This work will serve as a first step towards developing data-driven motion estimation algorithms for full body PET imaging. The innovation lies in the development of what is a data-driven solution to the problem of real time motion estimation.