Machine Learning with Scintillation Photon Counting Detectors to Advance PET Imaging Performance

点击翻译按钮获取中文摘要

Project Summary Clinical time-of-flight positron emission tomography (TOF-PET) systems capable of excellent coincidence time resolution (CTR) promise to drastically enhance effective 511 keV photon sensitivity. The ability to more precisely localize annihilation origins along system response lines constrains event data, providing improved signal-to- noise ratio (SNR) and reconstructed image quality by associating 511 keV photons more closely to their true origin. This SNR enhancement increases as CTR is improved, and a major goal of ongoing PET instrumentation research and development is to push system CTR ≤100 ps full-width-at-half-maximum (FWHM). At this level of performance, events are constrained ≤1.5 cm, providing a ≥five-fold increase in SNR relative to a system with no TOF capability. Advanced systems capable of ≤100 ps FWHM CTR would effectively more than double or quadruple the effective 511 keV system sensitivity, in comparison to state-of-the-art, clinical TOF-PET systems (250-400 ps FWHM CTR). Thus, advancing CTR is also a pathway for greatly improved system sensitivity without increasing detection volume and system material cost. Standard PET detectors comprising segmented arrays of high-aspect-ratio scintillation crystal elements and aggressive electronic signal multiplexing cannot achieve this level of performance and are ultimately limited by poor light collection efficiency, depth-dependent scintillation photon transit time jitter seen by the photodetector, and poor electronic SNR for optimal discriminator time pickoff and 511 keV photon time of interaction estimation. To address this, we are developing a new detector readout concept for monolithic scintillation detectors which allows scintillation photons arriving at each photosensor pixel to be counted and directly digitized. The spatiotemporal arrival time of scintillation photons in monolithic detectors intrinsically carries all information on 511 keV photon energy, three-dimensional (3D) position and time of interaction, and 3D position of interaction dependent scintillation photon transit skew. [Thus, this new detector readout concept’s ability to directly digitize the temporal scintillation light maps on photosensor arrays coupled to monolithic scintillators offers a unique opportunity for machine learning (ML) techniques to extract 3D positioning and time of interaction estimators in large area, thick (high 511 keV photon detection efficiency) detector modules that are at the statistical limit of performance. We will leverage this new advancement to investigate the performance of ML applied to the digitized photon data streams from a prototype detector module to demonstrate high resolution, three-dimensional positioning capabilities and CTR in a design that also makes no sacrifices on detection efficiency. The proposed PET detector technology can have a significant impact on quantitative PET imaging. The image SNR enabled by the significant boost in effective sensitivity can be employed to substantially reduce tracer dose and shorten scan time/increase patient throughput, or to better visualize and quantify smaller lesions/features in the presence of significant background, which are important features that can make PET more practical and accurate, as well as help to expand its roles in patient management.]