Self-supervised signal denoising in magnetic particle imaging

Various noises restrict magnetic particle imaging (MPI) to achieve higher resolution and sensitivity in practice. In this study, we proposed a self-supervised learning method to denoise MPI signals. The deep learning-based architecture consisted with four encoder’s blocks (EcBs) and four decoder’s blocks (DcBs). This model was trained with limited data of MPI magnetization signals to efficiently suppress noise related features by directly learning from the noisy signals. Simulated experiments showed that the self-supervised method could reduce the noise interference in MPI signals and eventually improve image quality.