FPGA Neural Networks Implementation for Nuclear Pulses Parameters Estimation

Abstract

Nuclear pulses parameters estimation is needed in many nuclear applications. Its precision and performance requirements are very demanding, especially in PET applications. Quality of PET images depends on the energy and time resolution of gamma pulses detection. Neural networks estimators were analyzed in contrast with common methods. Two-layer feed-forward networks with three neurons in the hidden layer reached precision goal. The chosen estimators allowed the use of 40 MHz free running ADC obtaining precision of 1ns in the timestamp determination, exceeding coincidence detection requirements. An efficient VHDL implementation on an inexpensive Xilinx Spartan-3 FPGA was achieved that fulfill performance requirements, adding no dead time due to digital processing. The estimators and its FPGA implementations were verified on hardware and characterization were done using nuclear shaped pulses synthesized with an arbitrary function generator.