Deep Learning Reconstruction for Improved TOF PET Using Histo-Image Partitioning

点击翻译按钮获取中文摘要

Project Summary Clinical and research applications of the PET imaging are rapidly expanding from ever improving diagnostic and treatment assessment applications to guidance of personalized treatments, ultra-low dose imaging, and even interventional imaging procedures. Supporting these developments, reconstruction tools that are able to reliably handle both typical and (ultra-)low count situations, imperfect data, and data from specialized imaging geometries, with fast (near real-time) reconstruction performance are of crucial importance. The overall goal of this project is to develop and investigate robust and efficacious Deep Learning (DL) reconstruction approaches addressing these needs. A unique and innovative feature of the proposed approaches (compared to alternative DL applications) is the utilization of list-mode data histogrammed into a very efficient histo-image format. TOF data partitioned into the histo-image format are characterized by strong local properties, thus perfectly fitting convolutional neural network formalism and making DL training and reconstruction directly from realistic clinical data (in size and character) highly feasible and practical. The clinical utility of PET systems has significantly improved over the years thanks to advances in instrumentation, data corrections, and reconstruction approaches. Nevertheless, full utilization of their potential through robust and fast quantitative reconstruction remains a challenge especially for the cases of very low count data, such as in low-count temporal (motion and dynamic) frames, delayed studies, longitudinal low-dose studies, and studies using new isotopes with long half-life and low positron fraction rates (e.g. in 89Zr-labeled CAR-T cell imaging), as well as in specialized PET systems with partial angular coverage, for which exact, artifact-free, reconstruction does not exist. These are the situations for which the developed DL approaches promise great potential due to the demonstrated success of the DL networks to be trained for imperfect and very low count data without reliance on accurate data models. Furthermore, pre-trained networks can provide ultra- fast, near real-time, performance in practical use. Specific Aim 1 will develop tools for DL PET reconstruction using histo-image partitioning along with procedures for training of the proposed DL approaches, including novel approaches advancing the state-of-the- art of DL reconstruction directly from acquired PET data. Specific Aim 2 is directed towards study and evaluation of the performance of the investigated DL approaches for whole-body and long axial FOV scanner data for the wide range of counts from applications such as typical FDG, low dose, delayed, low activity isotope scans, and ultra-short frames in motion correction and dynamic studies. Specific Aim 3 will develop and apply motion correction protocols involving the proposed DL reconstruction tools and test and study their efficacy for clinically realistic situations involving non-rigid lung and heart motions. And finally, Specific Aim 4 is dedicated to an application and study of the developed DL approaches to specialized PET systems with partial angular coverage. 1