Main Article Content
Copyright (c) 2023 Zhengyao Peng, Lin Yin, Jie Tian, Yang Du
This work is licensed under a Creative Commons Attribution 4.0 International License.
Magnetic particle imaging (MPI) is a rapidly evolving tomography modality. It has been proven to be potential in many biomedical fields for its high sensitivity and high image depth, but its spatial resolution is far from satisfactory. Deep learning may improve the spatial resolution of MPI, but it is limited by the need for large amount of training data. In order to conquer the above challenge, we propose a new deep learning solution to improve the spatial resolution of MPI in this work. The key of our method is to use a single system matrix to acquire the training data. which are further utilized to drive the deep learning model to improve the MPI spatial resolution. Firstly, some high-resolution (HR) images are used as the real magnetic particle distribution. Then, these HR images are downsampled to obtain the low-resolution (LR) particle distribution. To the end, the signal can be acquired by multiplying the LR particle distribution maps and LR system matrix. Hence, we can acquire the paired training data including the HR particle distribution maps and the signal data obtained with LR images. The results show that the data obtained through our proposed method can drive the deep learning model to achieve better performance in term of the spatial resolution than the bicubic interpolation.