Under-sampled magnetic resonance image reconstruction of an anatomical structure based on a machine-learned image reconstruction model

Patent metadata

Field	Value
Publication number	US-12482149-B2
Application number	US-202318130150-A
Country	US
Kind code	B2
Filing date	Apr 3, 2023
Priority date	Apr 3, 2023
Publication date	Nov 25, 2025
Grant date	Nov 25, 2025

Abstract

Disclosed herein are systems, methods, and instrumentalities associated with magnetic resonance (MR) image reconstruction. An under-sampled MR image may be reconstructed through an iterative process (e.g., over multiple iterations) based on a machine-learning (ML) model. The ML model may be obtained through a reinforcement learning process during which the ML model may be used to predict a correction to an input MR image of at least one of the multiple iterations, apply the correction to the input MR image to obtain a reconstructed MR image, determine a reward for the ML model based on the reconstructed MR image, and adjust the parameters of the ML model based on the reward. The reward may be determined using a pre-trained reward neural network and the ML model may also be pre-trained in a supervised manner before being refined through the reinforcement learning process.

First claim

What is claimed is:

1 . An apparatus, comprising:
at least one processor configured to:
obtain an under-sampled magnetic resonance (MR) image of an anatomical structure; and
reconstruct the under-sampled MR image of the anatomical structure through multiple iterations based on a machine-learned (ML) image reconstruction model, wherein the ML image reconstruction model is learned through a training process and during the training process:
the ML image reconstruction model is used to predict a correction to an input MR image obtained during at least one of the multiple iterations and generate a reconstructed MR image by applying the correction to the input MR image;
an ML reward model is used to determine a reward for the reconstructed MR image generated using the ML image reconstruction model; and
parameters of the ML image reconstruction model are adjusted based on the reward determined by the ML reward model, wherein, prior to being used in the training process of the ML image reconstruction model, the ML reward model is pre-trained for predicting a quality of an MR image and generating an evaluation for the MR image based on the predicted quality, the pre-training of the ML reward model is conducted based at least on a first MR training image, a second MR training image, and an indication that the second MR training image has a higher quality than the first MR training image, and the ML reward model is used during the pre-training to extract respective features from the first MR training image and the second MR training image and predict a quality of the first MR training image based on a difference between the respective features extracted from the first MR training image and the second MR training image.

2 . The apparatus of claim 1 , wherein, prior to being trained through the training process, the ML image reconstruction model is pre-trained based on under-sampled MR training images and corresponding fully-sampled MR images, and wherein initial parameters for the ML image reconstruction model are obtained based on the pre-training of the ML image reconstruction model.

3 . The apparatus of claim 1 , wherein, prior to being trained through the training process, the ML image reconstruction model is assigned random initial parameters.

4 . The apparatus of claim 1 , wherein the indication is received based on a human evaluation of the first MR training image and the second MR training image.

5 . The apparatus of claim 1 , wherein the input MR image obtained during the at least one of the multiple iterations is an output of a preceding iteration of the multiple iterations.

6 . The apparatus of claim 1 , wherein the ML image reconstruction model is trained to learn, through the training process, a probability distribution of the correction applied to the input MR image, and wherein the at least one processor is configured to reconstruct the under-sampled MR image based on a sample mean and a sample variance drawn from the probability distribution.

7 . The apparatus of claim 1 , wherein the ML image reconstruction model is trained to learn, through the training process, a mean of the correction applied to the input MR image, and wherein the at least one processor is configured to reconstruct the under-sampled MR image based at least on the mean of the correction.

8 . The apparatus of claim 7 , wherein the at least one processor is configured to reconstruct the under-sampled MR image further based on a constant variance applicable to the mean.

9 . The apparatus of claim 1 , wherein the correction comprises a map that includes a plurality of correction values to be applied to respective pixels of the input MR image.

10 . The apparatus of claim 1 , wherein the ML image reconstruction model is implemented via a recurrent neural network or a cascaded neural network.

11 . A method for medical image processing, the method comprising:
obtaining an under-sampled magnetic resonance (MR) image of an anatomical structure; and reconstructing the under-sampled MR image of the anatomical structure through multiple iterations based on a machine-learned (ML) image reconstruction model, wherein the ML image reconstruction model is learned through a training process and during the training process:
the ML image reconstruction model is used to predict a correction to an input MR image obtained during at least one of the multiple iterations and generate a reconstructed MR image by applying the correction to the input MR image;
an ML reward model is used to determine a reward for the reconstructed MR image generated using the ML image reconstruction model; and
parameters of the ML image reconstruction model are adjusted based on the reward determined by the ML reward model, wherein, prior to being used in the training process of the ML image reconstruction model, the ML reward model is pre-trained for predicting a quality of an MR image and generating an evaluation for the MR image based on the predicted quality, the pre-training of the ML reward model is conducted based at least on a first MR training image, a second MR training image, and an indication that the second MR training image has a higher quality than the first MR training image, and the ML reward model is used during the pre-training to extract respective features from the first MR training image and the second MR training image and predict a quality of the first MR training image based on a difference between the respective features extracted from the first MR training image and the second MR training image.

12 . The method of claim 11 , wherein, prior to being trained through the training process, the ML image reconstruction model is pre-trained based on under-sampled MR training images and corresponding fully-sampled MR images, and wherein initial parameters for the ML image reconstruction model are obtained based on the pre-training of the ML image reconstruction model.

13 . The method of claim 11 , wherein the indication is received as a part of a human evaluation of the first MR training image and the second MR training image.

14 . The method of claim 11 , wherein the ML image reconstruction model is trained to learn, through the training process, a probability distribution of the correction applied to the input MR image, and wherein the under-sampled MR image is reconstructed based on a sample mean and a sample variance drawn from the probability distribution.

15 . The method of claim 11 , wherein the ML image reconstruction model is trained to learn, through the training process, a mean of the correction applied to the input MR image, and wherein the under-sampled MR image is reconstructed based at least on the mean of the correction.

16 . A method for training a machine-learning (ML) image reconstruction model to reconstruct under-sampled magnetic resonance (MR) images, the method comprising:
predicting, during at least one of multiple iterations, a correction to an input magnetic resonance (MR) image based on present parameters of the ML image reconstruction model; applying the predicted correction to the input MR image to obtain a reconstructed MR image; determining, based on an ML reward model, a reward for the reconstructed MR image; and adjusting the present parameters of the ML image reconstruction model based on the determined reward, wherein, prior to being used in the training of the ML image reconstruction model, the ML reward model is pre-trained for predicting a quality of an MR image and generating an evaluation for the MR image based on the predicted quality, the pre-training of the ML reward model is

Assignees

• Shanghai United Imaging Intelligence Co Ltd

Inventors

• Sun Shanhui
• Chen Zhang
• Chen Xiao
• Liu Yikang
• Chen Terrence

Classifications

• G06T12/10 Primary

  Image preprocessing, e.g. calibration, positioning of sources or scatter correction · CPC title

• G06T2210/41

  Medical · CPC title

• G06T2211/424

  Iterative · CPC title

• G06T11/005 Primary

  Physics · mapped topic