Method and system for quantitative susceptibility mapping reconstruction in low-field magnetic resonance imaging (MRI) system

A system for Quantitative Susceptibility Mapping (QSM) reconstruction, comprising:

• one or more hardware processors;
• a communication interface;
• and
• memory storing a plurality of instructions, which when executed, cause the one or more hardware processors to:
• obtain at least one low-field measurement of a subject as input, wherein the at least one low-field measurement refers to low-field MRI imaging of the subject;
• enhance obtained at least one low-field measurement to generate an enhanced low-field measurement, using at least one pre-processing technique;
• construct a magnetic resonance (MR) image by processing the enhanced low-field measurement;
• remove perturbations from the MR image by performing forward operator drift correction;
• and
• reconstruct QSM image from the MR image, by iteratively performing:
• comparing a reconstructed QSM image obtained in each iteration with a reference image, to obtain a feedback representing extent of similarity of the reconstructed QSM image with the reference image, wherein the reference image is obtained by inverting the low-field measurement, further wherein the reconstructed QSM image is obtained based on an image prior enforced by an unsupervised CNN, and wherein the feedback is used as a loss function to update weights of the unsupervised CNN;
• refining the reconstructed QSM image from previous iteration in subsequent iteration, based on the obtained feedback;
• computing a similarity function at each iteration, wherein the similarity function at each iteration represents extent of similarity of the reconstructed QSM image with a structural image obtained from the subject;
• and
• terminating the iterative reconstruction when the extent of similarity represented by the similarity function exceeds a threshold,
• wherein the one or more hardware processors are configured to use the loss function to address non-uniformity in input data, wherein the loss function is represented using an equation
• ∥N(FHDFfθi(z)−ϕ)∥22+α∥fθi(z)−χi∥22
• wherein N is a noise matrix which accounts for non-uniformity, F and FH are Fourier transform and it's Hermetian, D is a dipole kernel, fθ(⋅) is a deep neural network parameterized by θ, ϕ is a MR phase, and χ is a susceptibility map.

One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

• obtaining at least one low-field measurement of a subject as input, wherein the at least one low-field measurement refers to low-field MRI imaging of the subject;
• enhancing obtained at least one low-field measurement to generate an enhanced low-field measurement, using at least one pre-processing technique;
• constructing a magnetic resonance (MR) image by processing the enhanced low-field measurement;
• removing perturbations from the MR image by performing forward operator drift correction;
• and
• reconstructing QSM image from the MR image, comprising, iteratively performing:
• comparing a reconstructed QSM image obtained in each iteration with a reference image to obtain a feedback representing extent of similarity of the reconstructed QSM image with the reference image, wherein the reference image is obtained by inverting the low-field measurement, further wherein the reconstructed QSM image is obtained based on an image prior enforced by an unsupervised CNN, and wherein the feedback is used as a loss function to update weights of the unsupervised CNN;
• refining the reconstructed QSM image from previous iteration in subsequent iteration, based on the obtained feedback;
• computing a similarity function at each iteration, wherein the similarity function at each iteration represents extent of similarity of the reconstructed QSM image with a structural image obtained from the subject;
• and
• terminating the iterative reconstruction when the extent of similarity represented by the similarity function exceeds a threshold,
• wherein the loss function is used to address non-uniformity in input data, wherein the loss function is represented using an equation
• ∥N(FHDFfθi(z)−ϕ)∥22+α∥fθi(z)−χi∥22
• wherein, N is a noise matrix which accounts for non-uniformity, F and FH are Fourier transform and it's Hermetian, D is a dipole kernel, fθ(⋅) is a deep neural network parameterized by θ, ϕ is a MR phase, and χ is a susceptibility map.