CSA-FCN: Channel- and Spatial-Gated Attention Mechanism Based Fully Complex-Valued Neural Network for System Matrix Calibration in Magnetic Particle Imaging

Abstract

Magnetic particle imaging (MPI) is an emerging medical imaging technique that visualizes the spatial distribution of magnetic nanoparticles (MNPs). The system matrix (SM)-based reconstruction is enable to sensitively account for various system imperfections and offers high-fidelity volume images. Yet, the re-calibration of SMs is time-consuming when the imaging mode changes. Here, through adequately analyzing the properties of SMs, a channel- and spatial- gated attention mechanism based fully complex-valued neural network (CSA-FCN) was introduced for SM calibration in MPI. Specifically, a complex-valued constraint model for SM calibration is designed to focus on the complex-valued property of SM samples. Firstly, complex-valued convolution neural network (C-CNN) is leveraged to coarsely extract complex-valued features of the SMs. Additionally, in complex-valued domain, the channel- and spatial-based gated attention mechanisms are constructed to enhance features with lightweight advantage, named C-SEM and C-SAM respectively. C-SEM induces the network to suppress the noise expression at channel-level. C-SAM improves the network context sensitivity at spatial-level. Ultimately, aggregate the features at each level as global embedding representation, and calibrating the SM form local- to full-size through a pre-constructed consistency reconstruction layer. Analysis and experiments indicate that CSA-FCN significantly improves the efficiency of SM calibration and has excellent robustness against to different imaging modes.