An Analog Neural Network ASIC for In-Sensor Processing of Radiation Detector Signals

Machine learning accelerators are increasingly used for fast, energy-efficient data processing, especially for large parallel computations, with analog in-memory computing offering higher computational density compared to digital platforms. This paper presents an analog neural network ASIC in 0.35 µm CMOS for in-sensor gamma-ray interaction localization in monolithic scintillators for nuclear medical imaging applications, such as PET and SPECT. Featuring 64 inputs, two hidden layers of 20 neurons, and two outputs, the ASIC processes photodetector signals in the charge domain, reducing data transmission and digitization, costs, latency, and power consumption, being particularly beneficial for large field-of-view detectors like total-body PET scanners. Experimental results demonstrate a 0.15 mm 2D FWHM spatial resolution and 47 GOP/J energy efficiency on simulated data. Despite circuit non-idealities impacting accuracy, the results support the integration of the ASIC with photodetector front-ends, and an improved architecture is proposed to address prototype limitations.