Improvement of the Spatial Resolution in Multi-Pinhole SPECT System with Deep Learning

Abstract

In a pinhole SPECT system, degradation of the spatial resolution occurs in reconstructed images depending on the pinhole aperture. Conventionally, blurring caused by such pinhole apertures has been corrected by obtaining a system matrix based on the detection probability or by increasing the number of projection rays in a ray driven method. However, the former has the drawback that the computational load increases when the pixel size is small or the number of pinholes increases, while the latter has the problem that the spatial resolution cannot be sufficiently improved when the pinhole diameter increases. Therefore, in this study, we proposed a method to reduce the deterioration of the spatial resolution due to pinholes using deep learning. Specifically, we implemented convolutional neural networks (U-net and U-net++) that learns projection data with reduced spatial resolution as input data and ideal infinitesimal pinhole projection data as training data to suppress the degradation of the spatial resolution. Simulation results showed that the peak signal-to-noise ratios of the corrected image with the U-net or U-net++ based method were 17.48 and 17.92 dB, respectively, and that with the conventional 21-rays method was 16.82 dB. It was clarified that the proposed methods can improve the spatial resolution compared with the conventional method.