A physics-informed deep learning framework in multi-color magnetic particle signal analysis

Abstract

Magnetic particle imaging (MPI) is an emerging and highly sensitive imaging method. Multi-color MPI allows simultaneous identification of different materials. Obtaining precise relaxation time is one of the key challenges in achieving multi-colored MPI. In this paper, we propose a physical information based deep learning framework to accurately decompose the mixed signal into the original independent relaxation signals. By transforming the Debye relaxation model into a differential loss function, our network is able to efficiently utilize physical prior information. In simulation experiments with different signal-to-noise ratios and different signal counts, our method shows better performance than the PDCO algorithm. The imaging effect of our algorithm and PDCO algorithm in the presence of multiple materials was evaluated by three-color imaging simulation experiment. In addition, spectral imaging of a digital vascular phantom was simulated by combining a field-free point with homogeneous pulsed excitation. In vascular phantom simulation experiment, our method images blood vessels, metal guidewires, and stents in a single imaging process, showing excellent application potential in cardiac stent surgery.