MRI Image Enhancement Training With Cropped Point Spread Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for magnetic resonance image reconstruction require substantial memory due to the need for full-sized training images, which is exacerbated by undersampled data acquisition, making it difficult to train efficiently.
Innovation Solution
The method involves cropping the point spread function in position space and using its Fourier transform to reduce the size of training images, allowing for supervised training with smaller images that match the size of a predefined spatial region, thereby reducing memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-sized training images are used for supervised training of machine learning models, then the training accuracy and model performance are improved, but the memory requirements and computational resources increase substantially
Solution Approach 1:
The patent divides the full-sized training image into multiple smaller patches or regions. Each patch is processed independently through the machine learning model during training, allowing the system to learn from local features while reducing memory consumption. The model processes these segmented regions sequentially rather than loading the entire large image into memory at once.
Solution Approach 2:
The patent transforms the problem from processing 2D full-sized images to processing a collection of smaller 2D patches, effectively adding a third dimension (the patch grid structure). This dimensional transformation allows the training process to handle large field-of-view requirements while maintaining reduced memory footprint by processing patches in a structured sequence.
2Productivity
If undersampled data acquisition is used to reduce scan time, then the productivity is improved, but the image quality and reconstruction accuracy deteriorate
Solution Approach 1:
The patent applies preliminary processing steps to the undersampled k-space data before reconstruction, including noise filtering and artifact suppression techniques. These preliminary actions prepare the data in a way that minimizes the negative effects of undersampling, allowing the subsequent machine learning model to recover image quality more effectively from the degraded input.
Solution Approach 2:
The patent introduces an intermediate processing stage between data acquisition and final reconstruction, where a preliminary reconstructed image or processed k-space data serves as an intermediary. This intermediary representation contains partially recovered information that the machine learning model then refines, effectively using the intermediary to bridge the gap between undersampled input and high-quality output.
3Measurement precision
If complex model architectures are used for deep learning reconstruction, then the reconstruction accuracy is improved, but the training time and computational complexity increase
Solution Approach 1:
The patent segments the complex model architecture into multiple simpler sub-networks or processing stages. Each sub-network performs a specific function (e.g., denoising, deblurring, artifact removal) in sequence, allowing the overall complex reconstruction task to be broken down into manageable components that can be trained more efficiently with reduced computational overhead at each stage.
Solution Approach 2:
The patent applies partial processing through the machine learning model, where not all patches or regions require the full computational power of the complete model architecture. Some regions may use simplified processing paths or fewer model layers, reducing overall training time while maintaining adequate reconstruction accuracy for the entire image.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach significantly reduces memory and runtime requirements for training machine learning models, enabling efficient training and reconstruction of magnetic resonance images even with undersampled data.
Implementation Method 1
the k-space and the position space are related to each other via Fourier transformation
Data Source
Figure 1~2
Figure 3~5
Figure 6~7
AI summary
For training a machine learning model, MLM, for image enhancement for use in MR image a point spread function (24) for undersampled MR data acquisition is received. A cropped point spread function (25) is determined, which is given by the point spread function (24) within a predefined spatial region (26). At least one training MR dataset corresponding to at least one coil channel is received and a ground truth reconstructed MR image corresponding to the at least one training MR dataset is received. The MLM is trained in a supervised manner depending on the at least one training MR dataset, on the ground truth reconstructed MR image, and on a Fourier transform of the cropped point spread function (25).