MRI Image Reconstruction Using Joint Machine Learning Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current MRI image processing methods often result in mutually independent reconstructions of different pulse sequences, leading to suboptimal image quality and increased scan times due to anisotropic resolution and lack of shared anatomical information.

Innovation Solution

A computer-implemented method using a trained machine learning processing unit, such as a U-net, to simultaneously reconstruct multiple MRI pulse sequences, leveraging supplemental data like B0-maps, phase encoding directions, and distortion information to enhance image quality and resolution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If multiple MRI pulse sequences are reconstructed independently, then the reconstruction process is simple and fast, but image quality and spatial resolution are suboptimal

Engineering Contradiction:
Improveimage qualityVSAvoidreconstruction complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple independently reconstructed MRI pulse sequences into a single joint reconstruction process. By combining the sequences and their corresponding k-space data, the system shares anatomical information and edge features across different sequences, thereby improving image quality and spatial resolution while maintaining manageable computational complexity through efficient data integration.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The reconstruction system is designed to handle multiple types of MRI pulse sequences simultaneously using a unified approach. The method can process different sequence types (e.g., T1-weighted, T2-weighted, FLAIR) through the same joint reconstruction framework, allowing the system to leverage complementary information from various sequences to enhance overall image quality without requiring sequence-specific processing pipelines.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Manufacturing precision

If high resolution 3D imaging is achieved using 2D sequences repeated in different planes, then spatial resolution can be improved, but scan time increases significantly

Engineering Contradiction:
Improvespatial resolutionVSAvoidscan time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent combines data from multiple 2D pulse sequences acquired in different imaging planes into a single joint reconstruction. By merging the k-space data and leveraging the anatomical overlaps between sequences, the system achieves high-resolution 3D-like images without requiring repeated 2D acquisitions, thereby reducing scan time while maintaining spatial resolution.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The method transitions from separate 2D plane reconstructions to a unified 3D reconstruction by integrating data across multiple imaging planes. This dimensional integration allows the system to exploit correlations between sequences and planes, achieving superior spatial resolution in three dimensions without the time penalty of acquiring complete 3D datasets through repeated 2D scans.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Productivity

If anisotropic resolution is used to reduce scan time, then scanning becomes faster, but image quality and signal-to-noise ratio deteriorate

Engineering Contradiction:
Improvescanning speedVSAvoidimage quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent merges multiple pulse sequences with anisotropic resolution into a joint reconstruction framework. By combining the sequences and their complementary information, the system recovers fine details and improves signal-to-noise ratio that would be lost in individual anisotropic acquisitions, thereby achieving high image quality without requiring isotropic sampling in all directions.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The method changes the resolution parameters dynamically during reconstruction by leveraging correlations between sequences. Instead of uniformly applying high resolution across all sequences (which would increase scan time), the system adapts resolution allocation based on sequence-specific characteristics and anatomical importance, maintaining high image quality while preserving the time efficiency of anisotropic sampling.

Inventive Principle:
Principle #35Parameter changes

4Reliability

If segmentation algorithms use only single pulse sequence data, then the processing is simple and fast, but segmentation reliability is reduced

Engineering Contradiction:
Improvesegmentation reliabilityVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges segmentation information from multiple pulse sequences by performing joint reconstruction first, which integrates anatomical features and edges across sequences. The resulting combined image data provides more robust and reliable segmentation results because it incorporates complementary contrast information from different sequences, improving the accuracy of organ, tumor, and structure contours despite increased processing complexity.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3798662B1Trained image processing for data sets of spin echo sequences
Publication Date: 2025.04.16 SIEMENS HEALTHINEERS AG
  • EP3798662B1 patent drawingFigure 1~4
  • EP3798662B1 patent drawingFigure 5~7
  • EP3798662B1 patent drawing

AI summary

The present invention relates to a computer-implemented method of training a machine learning based processing unit and a computer-implemented method of image processing for Magnetic Resonance Imaging (MRI) as well as respective computer programs, computer-readable media, data processing systems and an MRI system. The processing unit is trained to derive image data from signal data sets of multiple spin echo sequences.