Spatial signal distribution learning for high-resolution 3D system matrix calibration in magnetic particle imaging.

OBJECTIVE: Three-dimensional (3D) high-resolution system matrices (HR-SMs) are essential for high-quality image reconstruction in magnetic particle imaging (MPI), but obtaining HR-SMs is time-consuming and costly. This study aims to develop a learning-based 3D SM calibration method to reduce the calibration workload and maintain reconstruction accuracy. APPROACH: We propose a spatial signal distribution learning method for fast calibration of 3D HR-SMs. Specifically, a channel-decoupled multi-path state space model (MPC-SSM) is designed. This method flattens the 3D SM into multiple complementary spatial sequences, and uses different traversal paths to capture anisotropy and long-range spatial dependencies. To improve efficiency, feature channels are divided into disjoint groups and assigned to path-specific state space models (SSMs), achieving efficient multi-path sequence modeling while reducing computational overhead. MAIN RESULTS: We evaluated this method on both simulated and real MPI datasets (including OpenMPI). The results show that under 2× and 4× upsampling, the normalized reconstruction error of MPC-SSM is lower than that of existing interpolation and deep learning methods, and the quality of downstream image reconstruction is improved. SIGNIFICANCE: This work provides a scalable and practical solution for 3D HR-SM calibration and provides a general modeling strategy for structured 3D medical data.