Speaker
Description
Neural networks (NNs) are promising solutions for achieving fast and efficient reconstruction of the position of interaction in detectors, usually combining accuracy with ultra-low power operation. In scintillator-based PET and SPECT gamma cameras, embedding NN processing early in the acquisition chain can simplify the system design, reducing throughput, power consumption and scanner cost. NNs are mostly implemented in PCs, GUIs and FPGAs. In this work, we report the development of an Analog Neural Network ASIC (ANNA), which exploits the analog nature of detector signals and implements on-chip a NN without prior digitalization of signals. ANNA features a fully analog charge domain processing architecture enabling the implementation of configurable feedforward fully connected analog NN. Each neuron consists of an efficient multiplication and summing unit. The maximum supported complexity is 70 analog inputs with 5 weighted layers, up to 32 neurons per layer, and 32 analog outputs. Weights are quantized over 63 analog levels, with programmable non-linear activation functions. Preliminary measurements performed on the ASIC validate the correct operation of the proposed architecture and demonstrate the feasibility of fully analog gamma-ray position reconstruction on monolithic scintillators. Thanks to its programmability, ANNA employment is not limited to emission tomography, but can be tailored to other scenarios in processing signals of ”intelligent” detectors in fundamental and applied physics. We will discuss, in particular, the perspective of application of ANNA to reconstruct particle interaction in resistive LGADs (Low Gain Avalanche Diodes) towards the development of new compact and low-power high-energy physics trackers.