High quality low-dose SPECT reconstruction using CGAN-based transformer network with geometric tight framelet

Abstract

Single-photon emission computed tomography (SPECT) is a widely used diagnostic tool, but radioactive radiation during imaging poses potential health risks to subjects. Accurate low-dose single-photon emission computed tomography reconstruction is crucial in clinical applications of single-photon emission computed tomography. However, it remains a challenging problem due to the high noise and low spatial resolution of the low-dose reconstructed single-photon emission computed tomography images. The aim of the study is to develop a deep learning based framework for high quality low-dose single-photon emission computed tomography reconstruction. In the proposed framework, the conditional generative adversarial network (CGAN) was used as backbone structure and a Residual Attention CSwin Transformer (RACT) block was introduced as the basic building block for the generator of conditional generative adversarial network. The proposed residual attention CSwin transformer block has a dual-branch structure, which integrates the local modeling capability of CNN and the global dependency modeling capability of Transformer to improve the quality of single-photon emission computed tomography reconstructed images. More importantly, a novel loss term based on the geometric tight framelet (GTF) was designed to better suppress noise for the single-photon emission computed tomography reconstructed image while preserving the details of image to the greatest extent. Monte Carlo simulation software SIMIND was used to produce low-dose single-photon emission computed tomography images dataset to evaluate the performance of the proposed method. The simulation results showed that the proposed method can reduce more noise and preserve more details of various situations compared to several recent methods. To further validate the performance of the proposed method, we also verified the generalization ability of the proposed method, which is more adaptable to different noise level scenarios than other methods. Our results indicated that the proposed framework has the potential to reduce the tracer dose required for single-photon emission computed tomography imaging without compromising the quality of the reconstructed images.