On-Chip Analog Neural Network for Edge Computing in Radiation Detectors

This work presents an analog on-chip neural network for embedded processing of gamma rays in radiation detectors, featuring 64 inputs, two 20-neurons hidden layers, and two outputs. Leveraging in-sensor processing of analog signals coming from photodetectors permits to reduce the amount of data to transmit and digitize, as well as eliminating the need of FPGAs for signal processing, allowing for an easier scale up of complex multichannel systems. Fabricated in a 0.35 μm CMOS process node, this first prototype chip demonstrates an energy efficiency of 46.8 GOP/J. Its functionality is showcased through the localization of X and Y coordinates of gamma photons interacting in a scintillator crystal readout by a planar array of silicon photomultipliers (SiPMs), in the field of medical imaging applications based on emission tomography (such as PET and SPECT). The neural network’s weight reconfigurability broadens its applicability beyond gamma-ray detection, making it suitable for other edge-computing applications requiring a feedforward neural network architecture.