US-12620152-B2 - Iterative hierarchal network for regulating image reconstruction

Abstract

For reconstruction in sampling-based imaging, such as reconstruction in MR imaging, an iterative, multiple-mapping based hierarchal machine-learned network reconstruction may produce artifact corrected images based on under-sampled scans. Two or more mappings may be used to reduce the presence of artifacts, in some cases including localized low-noise-contribution artifacts, relative to reconstructions based on fully-sampled scans.

Inventors

• Antoine Edouard Adrien Cadiou
• Mahmoud Mostapha
• Simon Arberet
• Mariappan S. Nadar

Assignees

• Siemens Healthineers Ag

Dates

Publication Date

20260505

Application Date

20230217

Claims (20)

1. 1 . A method for reconstruction of an output image, the method including: performing, by an imaging system, a capture scan of an object; obtaining, via processing separate from reconstruction of the output image, a first mapping via a decomposition of reference data corresponding to the imaging system; obtaining, via the processing separate from reconstruction of the output image, a second mapping via the decomposition of the reference data corresponding to the imaging system; and reconstructing the output image by: obtaining, at a first consistency stage implemented in a machine-learned network including iterative hierarchical convolutional networks, a base image corresponding to the first mapping and the capture scan; obtaining, at the first consistency stage, a secondary image corresponding to the second mapping and the capture scan; and generating an iteration of the output image by deriving the iteration of the output image based on a concatenation of the base image and the secondary image, where the iteration of the output image differs from the base image by at least an artifact-corrected pixel altered based on signal data present in the secondary image.
2. 2 . The method of claim 1 , wherein the first consistency stage includes a final consistency stage among the iterative hierarchical convolutional networks.
3. 3 . The method of claim 1 , wherein the first consistency stage includes an initial consistency stage among the iterative hierarchical convolutional networks.
4. 4 . The method of claim 1 , wherein obtaining the base image at the first consistency stage includes obtaining the base image from a previous iteration of the iterative hierarchical convolutional networks.
5. 5 . The method of claim 1 , further including implementing the first consistency stage within a cascade of the iterative hierarchical convolutional networks.
6. 6 . The method of claim 1 , wherein: the first mapping corresponds to a first eigenvector derived from the decomposition of the reference data; and the second mapping corresponds to a second eigenvector derived from the decomposition of the reference data.
7. 7 . The method of claim 1 , wherein obtaining, at the first consistency stage, the first mapping and the second mapping includes receiving the first mapping and the second mapping in a cascade from one or more mapping estimation networks within the machine-learned network.
8. 8 . The method of claim 1 , wherein reconstructing the output image further includes obtaining, at the first consistency stage, one or more tertiary images each corresponding to the capture scan and a respective one of one or more third mappings, the one or more third mappings each derived from the reference data.
9. 9 . The method of claim 8 , wherein a total number of mappings used to reconstruct the output image is determined via adjustment of a mapping number parameter of the imaging system.
10. 10 . The method of claim 1 , further including obtaining the reference data by performing, by the imaging system, a reference scan, the reference scan being performed at a higher sample density than the capture scan.
11. 11 . The method of claim 1 , wherein the first mapping and the second mapping include coil sensitivity mappings.
12. 12 . The method of claim 1 , wherein the imaging system includes a magnetic resonance imaging (MRI) system.
13. 13 . A product including: non-transitory machine-readable media; and instructions stored on the machine-readable media, the instructions configured to, when executed, cause a processor to: obtain, from an imaging system, a capture scan of an object; obtain, via processing separate from reconstruction of an output image, a first mapping via a decomposition of reference data corresponding to the imaging system; obtain, via the processing separate from reconstruction of the output image, a second mapping via the decomposition of the reference data corresponding to the imaging system; and reconstruct the output image by: obtaining, at a first consistency stage implemented in a machine-learned network including iterative hierarchical convolutional networks, a base image corresponding to the first mapping and the capture scan; obtaining, at the first consistency stage, a secondary image corresponding to the second mapping and the capture scan; and generating an iteration of the output image by deriving the iteration of the output image based on a concatenation of the base image and the secondary image, where the iteration of the output image differs from the base image by at least an artifact-corrected pixel altered based on signal data present in the secondary image.
14. 14 . The product of claim 13 , wherein the instructions are further configured to cause the processor to obtain the output image further by obtaining, at the first consistency stage, one or more tertiary images each corresponding to the capture scan and a respective one of one or more third mappings, the one or more third mappings each derived from the reference data.
15. 15 . The product of claim 14 , wherein a total number of mappings used to reconstruct the output image is determined via adjustment of a mapping number parameter of the imaging system.
16. 16 . The product of claim 13 , wherein the instructions are further configured to obtain the reference data by performing, by the imaging system, a reference scan, the reference scan being performed at a higher sample density than the capture scan.
17. 17 . An imaging system including: imaging scanner configured to perform a capture scan of an object; and reconstruction circuitry, in data communication with the imaging scanner, the reconstruction circuitry configured to reconstruct an output image by: obtaining, via processing separate from reconstruction of the output image, a first mapping via a decomposition of reference data corresponding to the imaging system; obtaining, via the processing separate from reconstruction of the output image, a second mapping via the decomposition of the reference data corresponding to the imaging system; obtaining, at a first consistency stage implemented in a machine-learned network including iterative hierarchical convolutional networks, a base image generated using the first mapping and the capture scan as inputs; obtaining, at the first consistency stage, a secondary image generated using the second mapping and the capture scan as inputs; and generating an iteration of the output image by deriving the iteration of the output image based on a concatenation of the base image and the secondary image, where the iteration of the output image differs from the base image by at least an artifact-corrected pixel altered based on signal data present in the secondary image.
18. 18 . The imaging system of claim 17 , wherein the reconstruction circuitry is configured to implement the first consistency stage within a cascade of the iterative hierarchical convolutional networks.
19. 19 . The imaging system of claim 17 , wherein: the first mapping corresponds to a first eigenvector derived from the decomposition of the reference data; and the second mapping corresponds to a second eigenvector derived from the decomposition of the reference data.
20. 20 . The imaging system of claim 17 , wherein the reconstruction circuitry is configured to obtain the first mapping and the second mapping by receiving the first mapping and the second mapping in a cascade from one or more mapping estimation networks within the machine-learned network.

Description

PRIORITY This application claims priority to U.S. Provisional Application No. 63/374,034 filed Aug. 31, 2022, and titled Deep Unrolled Reconstruction Network for Highly Undersampled MRI Images Using Multiple Sets of Coil Sensitivity Maps, which is incorporated by reference herein in its entirety. FIELD This disclosure relates to image reconstruction, such as reconstruction in magnetic resonance (MR) imaging. BACKGROUND Sampling-based imaging, such as, various forms of medical imaging, magnetic resonance (MR), computed tomography (CT), positron emission tomography (PET), and/or single photon emission computed tomography (SPECT), use reconstruction to estimate an image or real-space object from measurements. These scans may be time consuming. For example, numerous methods have been proposed to accelerate the MR scan. One acceleration method is the under-sampling reconstruction technique (e.g., MR parallel imaging with compressed sensing (CS)), where fewer samples are acquired in the MRI data space (k-space), and prior knowledge is used to restore the images in reconstruction. MR results obtained using compressed sensing reconstruction tend to show unfolding artifacts. An image regularizer is used in reconstruction to reduce these aliasing artifacts, but the regularizer adds computational burden. Deep learning (DL) techniques based on unfolding (unrolled) iterative reconstruction algorithms with learnable regularization improve the speed and the reconstruction quality compared to CS. Some DL-based image reconstruction methods are based on unrolled iterative algorithms where a data-consistency step alternates with a regularization network. In order to obtain good results, multiple unrolled iterations of reconstruction are performed. Computational time and memory requirements are directly proportional to the number of unrolled iterations. Deep learning models need to be fast and memory-efficient while also robust to variations in MRI intensities and contrasts originating from using different scanned organs, acquisition parameters, and image resolutions. Current MRI reconstruction schemes typically utilize image regularization deep learning networks in the form of an encoding-decoding structure such as different U-net architectures. Decreasing and increasing the resolution of the feature maps is effective for learning from heterogeneous datasets, but U-net architectures increase the overall size of the feature maps, resulting in decreasing the receptive field and increasing the computational complexity. Designing robust deep learning image regularization networks is critical in constructing high-quality MRI from subsampled multi-coil data acquired with a wide range of varying MRI acquisition protocols and scanner models. Such networks would avoid MR reconstructions with degraded image quality and reduced clinical value. The encoding-decoding structure (e.g., different U-net architectures) is trained on large datasets that cover the expected MRI variability at test time. However, in practice, learning from such large datasets requires deep learning networks with enormous capacity, which increases the training time and increases their computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of an implementation of an MR system for sampling-based imaging using an iterative and/or hierarchal network for regulating; FIG. 2 shows an implementation of iteration in reconstruction; FIG. 3 is a block diagram of one implementation of an iterative network for regulating in a given iteration of reconstruction; FIG. 4 is a block diagram of one implementation of a hierarchal U-block network for regulating; FIG. 5 is a block diagram of one implementation of a U-block architecture used in a hierarchal U-block network; FIG. 6 is a flow chart diagram of one implementation of a method for machine training for regularizing in reconstruction; FIG. 7 is a flow chart diagram of one implementation of a method for reconstruction using a machine-learned network for regularization; FIG. 8 is a block diagram of one implementation of an iterative network for regulating in a given iteration of reconstruction, using multiple mappings; and FIG. 9 is a flow chart diagram of one implementation of a method for reconstruction using a machine-learned network for regularization using multiple mappings. FIG. 10 shows an example deepsets coil sensitivity mapping (CSM) computation system. FIG. 11 shows an example deepsets CSM estimation network. FIG. 12 shows an example deepsets CSM refinement network. FIG. 13 shows example end-to-end reconstruction logic. DETAILED DESCRIPTION By way of introduction, the various implementations described below include methods, systems, instructions, and computer readable media for reconstruction in sampling-based imaging, such as reconstruction in MR imaging and/or other medical imaging technology. An iterative, hierarchal network for regularization may decrease computational complexity. In v