A Sensor Network Architecture for Digital SiPM-Based PET Systems

Digital silicon photo-multipliers (SiPMs) have emerged in the recent past as a viable low cost alternative to photomultiplier tubes in positron emission tomography systems, providing multiple timestamps, energy and scintillation coordinates at high spatial granularity as well as MRI compatibility. The rich but large datasets generated by digital SiPM sensors have posed a data preprocessing and acquisition challenge, at the sensor, module and system level when a multitude of such sensors are to be used. In this paper, we present a sensor network-based approach for data acquisition, scalable to multiring configurations, whereby each module acts as an autonomous sensing and computing unit, capable of determining in real time basic information for each scintillation event and communicating it to its peers. The proposed architecture is equally applicable to modules based on analog SiPMs with local digitization. Coincidence detection can then take place in the ring itself, in a deferred and distributed manner to ensure scalability and allow to fully process only the fraction of the total events which corresponds to true coincidences. Simulations and experimental results show that it is indeed possible to handle the system level challenges associated with digital SiPMs at data rates compatible with realistic configurations, including event packet transfers and real-time coincidence detection, using Gb/s serial communication links for internode communication. The downside of the proposed architecture is represented by the need, at module level, for additional connectivity and processing power. We also address possible solutions for network-based clock synchronization, in particular a hybrid scheme, combining a hard-wired ring clock distribution network with a network-based clock offset estimator. The latter was tested in an 8-node system, performing synchronization in real time with a worst-case phase estimator stability (standard deviation of the clock phase offset estimation between two nodes) of about 160 ps, and substantial room for improvement (at least 5x according to simulations)