A Field Free Line 3D Reconstruction Model for Magnetic Particle Imaging for Improved Sensitivity, Resolution, and High Dynamic Range Imaging

TL;DR

This work tackles the challenge of reconstructing high-quality 3D MPI images from time-domain data by introducing a physics-based 3D field-free line (FFL) reconstruction model that integrates multi-angle projections with tomographic operators. A harmonic-domain compression stage is embedded in the forward model, dramatically reducing memory demands (≈99.9%) and enabling GPU-based volumetric reconstructions in minutes. The approach yields substantial gains in spatial resolution and detection sensitivity, reported as an ~11× improvement in iron-detection limits over conventional X-space CT in phantom studies, and significantly reduces background haze in in vivo images, improving visualization of low-intensity structures near bright organs. These advances provide a scalable, model-based MPI framework with practical implications for preclinical and potential clinical MPI imaging, especially in high-dynamic-range scenarios.

Abstract

Magnetic particle imaging (MPI) is a tracer-based imaging modality that detects superparamagnetic iron oxide nanoparticles in vivo, with applications in cancer cell tracking, lymph node mapping, and cell therapy monitoring. We introduce a new 3D image reconstruction framework for MPI data acquired using multi-angle field-free line (FFL) scans, demonstrating improvements in spatial resolution, quantitative accuracy, and high dynamic range performance over conventional sequential reconstruction pipelines. The framework is built by combining a physics-based FFL signal model with tomographic projection operators to form an efficient 3D forward operator, enabling the full dataset to be reconstructed jointly rather than as a series of independent 2D projections. A harmonic-domain compression step is incorporated naturally within this operator formulation, reducing memory overhead by over two orders of magnitude while preserving the structure and fidelity of the model, enabling volumetric reconstructions on standard desktop GPU hardware in only minutes. Phantom and in vivo results demonstrate substantially reduced background haze and improved visualization of low-intensity regions adjacent to bright structures, with an estimated $\sim$11$\times$ improvement in iron detection sensitivity relative to the conventional X-space CT approach. These advances enhance MPI image quality and quantitative reliability, supporting broader use of MPI in preclinical and future clinical imaging.

Figures (7)

• Figure 1: 3D volume renderings of in vivo MPI reconstructions comparing the proposed model-based 3D FFL reconstruction (left) to the conventional sequential X-space CT reconstruction (right). A cool-to-warm colormap is used, with low-intensity values shown in light blue and high-intensity values shown in red. The new model-based reconstruction exhibits improved spatial resolution and reduced background haze, particularly around high-signal regions such as the liver and tumor. This enhancement enables more accurate quantification of nearby low-intensity regions that are often obscured in the conventional reconstruction.
• Figure 2: Compression of the time domain signal. Top left: simulated time domain MPI signal data. Top right: spectrum magnitude of the MPI signal data. Bottom left: logarithm of the spectrum magnitude. Bottom right: final compressed data form used for the reconstruction. The compressed form of the data reduces to total data memory overhead by 99.90%.
• Figure 3: Diagram of the FFL geometry in 3D. At lines along the 20$^{\circ}$ FFL angle, the magnetic field is a constant. Therefore, when the transmit amplitude is orthogonal to this line, the full 3D MPI model can be constructed with a 3D to 2D projection operation followed by the 2D MPI model.
• Figure 4: Results on sensitivity series data comparing the 3D FFL model (top row) with the X-space CT projection reconstruction (bottom row). The images on the left show the results from the higher concentration samples (5-50 ug) and the right images show the results from the lower concentration data (0-1 ug). The left two columns show maximum intensity projections (MIPs), and the right images show a single slice of the 3D volume, where the location is indicated by the red dashed line in the middle column.
• Figure 5: Analysis of the detection limit with the new model vs. the X-space CT reconstruction using the sensitivity data sets. The detection limit for the new model was determined to be 0.23 ug, while for the X-space CT reconstruction the limit was 2.54 ug. This corresponds to an 11.2x improvement in the detection limit with the new model. Note that a higher drive amplitude of 25 mT would theoretically improve both detection limits by another factor of 25 (see footnote).