Multidimensional Information Fusion Learning for Real-time Signal Denoising in Magnetic Particle Imaging

Magnetic Particle Imaging (MPI) is a novel molecular imaging technique that enables highly sensitive visualization of the distribution of magnetic nanoparticles within biological tissues. However, dynamic random noises contained in MPI signal leads to artifacts. A common approach to improve signal-to-noise ratio (SNR) is averaging signals from multiple repeated measurements, significantly reducing real-time performance. To address this, we propose a deep learning method to remove random noise from short-time measurement signals, making them close to long-time measurement signals, thereby shortening the measurement time. We integrate the frequency indexes and coil channels of frequency components using a Transformer architecture, embedding it into the frequency domain signal to form multidimensional information. Subsequently, we introduce the row energy normalization method for signal preprocessing in model training. This approach makes the energy of the frequency components consistent, enhancing image reconstruction quality. Experiments on real-world datasets, including the OpenMPI dataset, the phantom dataset, and in vivo dataset, demonstrate our method can significantly eliminate random noise, improve the SNR of limited measurement signals, and obtain high-quality reconstructed images. This method can significantly improve the SNR on the phantom dataset from -24.1417 dB to 1.0184 dB, while shortening the acquisition time by 92.92%. Our method shortens acquisition time while ensuring imaging quality and robustness, enhancing the real-time imaging performance of MPI and providing a new approach for its biomedical applications.