PGNet: Projection generative network for sparse-view reconstruction of projection-based magnetic particle imaging

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Magnetic Particle Imaging (MPI) is a novel tomographic modality for imaging superparamagnetic iron oxide nanoparticles.
• Acquiring multiview projections for 3D MPI reconstruction is time-consuming, limiting temporal resolution.
• Sparse-view reconstruction techniques, successful in CT, have not been explored for MPI.

Purpose of the Study:

• To improve the 3D imaging temporal resolution of projection MPI.
• To address the challenge of time-consuming multiview projection acquisition in 3D MPI.
• To enable faster 3D MPI by utilizing sparse-view projections for reconstruction.

Main Methods:

• Proposed a novel deep learning method, the projection generative network (PGNet), to generate new projections from sparse-view data.
• PGNet incorporates an attention mechanism, adversarial training, and a fusion loss function.
• Developed a comprehensive dataset including simulations, phantom, and in vivo mouse data for training and validation.

Main Results:

• PGNet effectively generates novel projections, enabling artifact-free MPI tomographic imaging.
• The method significantly suppresses streaking artifacts compared to existing sparse-view techniques.
• Achieved a 6.6-fold improvement in 3D imaging temporal resolution for projection MPI.

Conclusions:

• A deep learning approach in the projection domain addresses sparse-view reconstruction challenges in MPI.
• The constructed sparse-dense dataset alleviates data scarcity issues in MPI reconstruction.
• The method reduces the need for extensive real projection acquisitions, preventing artifacts and showing potential for time-sensitive in vivo 3D MPI.