We present an analog ASIC implementing a neural network (NN) in charge domain for estimating the position of interaction of gamma rays in Anger Cameras. The multiplyand-accumulate (MAC) operations are performed by means of programmable switched-capacitor crossbar arrays. The analog NN consists of 64 input nodes receiving the SiPM signals, two hidden layers of 20 neurons each, and two outputs nodes corresponding to the inference of the x and y position of interaction. The network parameters (weights and biases) are obtained with an off-line training, using a reference dataset for which the 2D position of interaction of the events is known. To estimate the performance of the ASIC, which is currently in the final design phase and will be soon submitted for production, its positioning capabilities were tested using two different datasets. In both cases, the network parameters are quantized with 5-bit precision to emulate computations in a CMOS process. The first test was performed using the data obtained with a Monte Carlo simulation of a 51 mm×51 mm×10 mm LYSO scintillator read out by an 8×8 array of SiPMs: from high-level MATLAB simulations, consistent with ASIC topology and operations, the NN achieves a spatial resolution of 1.22 mm FWHM/1.66 mm MAE. The second test was performed using experimental data obtained with a PET detector based on 32 mm×32 mm×22 mm LYSO:Ce scintillator. For this detector, we report a comparison of the achieved spatial resolution with respect to a k nearest neighbour (k-NN) method, which shows that the NN achieves comparable results.

Towards an Analog Neural Network ASIC (ANNA) for Position Reconstruction in Anger Cameras

Di Giacomo, S.;Ronchi, M.;Borghi, G.;Carminati, M.;Fiorini, C. E.
2023-01-01

Abstract

We present an analog ASIC implementing a neural network (NN) in charge domain for estimating the position of interaction of gamma rays in Anger Cameras. The multiplyand-accumulate (MAC) operations are performed by means of programmable switched-capacitor crossbar arrays. The analog NN consists of 64 input nodes receiving the SiPM signals, two hidden layers of 20 neurons each, and two outputs nodes corresponding to the inference of the x and y position of interaction. The network parameters (weights and biases) are obtained with an off-line training, using a reference dataset for which the 2D position of interaction of the events is known. To estimate the performance of the ASIC, which is currently in the final design phase and will be soon submitted for production, its positioning capabilities were tested using two different datasets. In both cases, the network parameters are quantized with 5-bit precision to emulate computations in a CMOS process. The first test was performed using the data obtained with a Monte Carlo simulation of a 51 mm×51 mm×10 mm LYSO scintillator read out by an 8×8 array of SiPMs: from high-level MATLAB simulations, consistent with ASIC topology and operations, the NN achieves a spatial resolution of 1.22 mm FWHM/1.66 mm MAE. The second test was performed using experimental data obtained with a PET detector based on 32 mm×32 mm×22 mm LYSO:Ce scintillator. For this detector, we report a comparison of the achieved spatial resolution with respect to a k nearest neighbour (k-NN) method, which shows that the NN achieves comparable results.
2023
979-8-3503-3866-9
File in questo prodotto:
File Dimensione Formato  
Towards_an_Analog_Neural_Network_ASIC_ANNA_for_Position_Reconstruction_in_Anger_Cameras.pdf

Accesso riservato

: Publisher’s version
Dimensione 723.06 kB
Formato Adobe PDF
723.06 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1259699
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact