DeepSPIO: Super Paramagnetic Iron Oxide Particle Quantification Using Deep Learning in Magnetic Resonance Imaging.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Superparamagnetic iron oxide (SPIO) particles are valuable MRI contrast agents.
• Current quantification methods (relaxometry, inhomogeneity measurement) are unreliable at high SPIO concentrations.
• Phase-based quantification methods used in MRI are limited by concentration.

Purpose of the Study:

• To develop a novel Deep Learning (DL) method for accurate SPIO concentration distribution quantification.
• To address the limitations of existing phase-dependent MRI quantification techniques.
• To improve the reliability of SPIO quantification in MRI.

Main Methods:

• A novel DL network utilizing residual blocks and multiple decoders for improved gradient flow.
• A new MRI sequence, View Line, encoding field map information in image geometry.
• Wavelet decomposition of the concentration map, with each decoder predicting a different part.

Main Results:

• The DL model accurately quantified SPIO concentration distribution in both simulated and phantom data.
• The wavelet decomposition approach improved concentration estimation and accelerated model convergence.
• The method demonstrated reliable performance across various SPIO concentrations.

Conclusions:

• The proposed DL method offers an accurate and reliable approach for SPIO concentration quantification in MRI.
• The View Line sequence and the novel network architecture overcome previous limitations.
• This technique has the potential to enhance the utility of SPIO nanoparticles in MRI applications.