Towards Analog Neural Networks Integrated in Detectors Readout

The conceptual design of a novel ASIC in 0.35 mu m CMOS technology implementing in an analog way an artificial neural network (ANN) with 64 inputs, 2 20-neuron hidden layers and 2 outputs is presented. The ANN was optimized to address the application in medical imaging based on emission tomography, in particular the estimation of the position of absorption of a gamma-ray photon in a crystal scintillator readout by a planar array of silicon photomultipliers. The key advantage of this analog approach, as opposed to digital ones, is the possibility of integrating local processing with the detector front-end and, thus, significantly reduce the number of signals (from 64 to 2) that have to be transmitted and digitized. This would allow an easier scale up of scanners for molecular imaging and diagnostics with larger fields of view, i.e. larger amount of imaging modules and/or pixels per module. However, given the widespread use of ANNs in a broad variety of sensing contexts, we believe that such as concept could be extended to several applications beyond gamma cameras.