One of the key advances in modern positron emission tomography (PET) systems for clinical imaging is the use of time-of-flight (TOF) information. In cancer imaging TOF provides superior lesion detection and more accurate quantification that is crucial in measuring response to therapy, as well as in neurological and cardiovascular imaging applications. An additional benefit of TOF is the ability to lower the radiation dose or scan time without sacrificing image quality, important for patient safety and comfort. The magnitude of these clinical benefits is determined by the TOF resolution of the PET detectors, therefore the prospect of achieving unprecedented image quality and clinical imaging capabilities with superior TOF resolution has fueled significant research in developing detector technology for TOF-PET systems. However, these developments have been largely unaccompanied by advances in signal processing methods needed to extract TOF information from the detector?s electrical signals, with most detectors making use of crude analog algorithms that discard most of the useful timing information contained in the signals. Now, with the availability of low-cost fast waveform digitizers, there is an exciting opportunity to implement sophisticated digital signal processing algorithms to achieve superior TOF resolution. The main advantage of developing advanced signal processing algorithms is that it presents a cost-effective route to improved TOF resolution that is complementary to instrumentation innovations. In essence, the TOF gain comes for free; the detector signals already contain the information needed for better TOF resolution, it just needs to be used effectively. Here we propose to tailor deep learning techniques to estimate TOF from the detector signals. Deep learning with convolutional neural networks (CNNs) is a powerful approach to learn complex representations of input data that can be used for tasks such as classification and regression. CNNs are therefore very suitable for directly estimating TOF from the detector waveforms, since these waveforms are influenced by several complex and intertwined processes which are hard to accurately model. Furthermore, large amounts of ground truth training data are readily generated. We recently demonstrated the feasibility of CNN-based TOF estimation, and found up to 23% improvement in TOF resolution compared to standard signal processing methods. This proposal aims to optimize these methods to push the limits of achievable TOF resolution and develop methods for their practical implementation. First, we will develop CNN architectures and methods suitable for silicon photomultipliers (SiPMs) that are now used in modern TOF-PET systems. We will also optimize the digitizing parameters to make optimal use of CNNs for TOF estimation. Second, we will implement CNN-TOF methods in a modern commercial PET detector, including using a global CNN to simultaneously estimate TOF and the crystal-of-interaction from the detector waveforms, demonstrating the practical feasibility of using this promising deep learning method in next generation PET systems.
The use of time-of-flight information in positron emission tomography (PET) is a powerful way to increase clinical imaging capabilities of PET, and is now used in most modern systems. Improving time-of-flight performance promises to provide substantial benefits for lesion detection in cancer imaging, kinetic modeling with dynamic imaging, and the use of lower radiation dose for patient safety. Here we introduce deep learning signal processing methods to significantly improve the performance of time-of-flight PET detectors without requiring new materials, which thereby represents a cost effective route towards maximizing the use of the rich imaging signal provided by time-of-flight.
