MRI Reconstruction Using Deep Learning to Resolve Sub-sampling Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Magnetic Resonance Imaging (MRI) data acquisition is limited by the Nyquist sampling theorem, leading to long sampling times and potential motion artifacts, and sub-sampling techniques often result in aliasing artifacts during image reconstruction.
Innovation Solution
A method using a machine learning-based image reconstruction model that processes sub-sampled images from multiple MRI sequences with different sub-sampling rates to generate a full image, leveraging reference full images for improved quality and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sub-sampling technique is used to accelerate imaging process, then imaging speed is improved, but aliasing artifacts are generated during image reconstruction
Solution Approach 1:
A deep learning reconstruction model serves as an intermediary between the sub-sampled k-space data and the final image. The model learns the mapping relationship from subsampled to full images during training, and during inference, it reconstructs high-quality images from low-resolution subsampled data, effectively eliminating aliasing artifacts while maintaining accelerated imaging speed
Solution Approach 2:
The patent changes the reconstruction approach from traditional iterative methods to a data-driven deep learning model. The model parameters are trained to optimize the mapping from subsampled to full images, enabling high-quality reconstruction from highly subsampled data without the aliasing artifacts that plague conventional methods
2Manufacturing precision
If traditional MRI sampling is used to ensure image quality, then measurement precision is maintained, but scanning time increases
Solution Approach 1:
The patent applies partial sampling (sub-sampling) of k-space data to reduce acquisition time. Instead of collecting all k-space data according to Nyquist theorem, the system collects only a subset and uses a deep learning model to reconstruct the complete image, achieving both faster scanning and maintained image quality through intelligent reconstruction
3Productivity
If sub-sampling rate is increased to reduce scan time, then productivity is improved, but image quality deteriorates due to aliasing artifacts
Solution Approach 1:
The patent replaces traditional mechanical sampling methods with a computational deep learning approach. The deep learning model substitutes for the need to collect complete k-space data, allowing high sub-sampling rates (e.g., 4x or higher) while maintaining image quality through learned reconstruction rather than conventional signal processing
Data Source
AI summary
A method for MRI reconstruction is provided. The method may include obtaining a plurality of sub-sampled images of a subject. The plurality of sub-sampled images may include a first sub-sampled image of the subject and one or more second sub-sampled images of the subject. The first sub-sampled image may be generated using a first MRI sequence and a first sub-sampling rate. Each of the one or more second sub-sampled images may be generated using a second MRI sequence and a second sub-sampling rate. The second sub-sampling rate may be smaller than the first sub-sampling rate. The method may include obtaining an image reconstruction model having been trained according to a machine learning technique. The method may further include generating a first full image of the subject corresponding to the first MRI sequence based on the first sub-sampled image, the one or more second sub-sampled images, and the image reconstruction model.