MRI Denoising via Multi-Coil Surrogate Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for denoising MRI images using supervised learning face challenges in obtaining true images for training, as they require either the true image or a label image with similar noise characteristics, which can be difficult or impossible to prepare, hindering efficient denoising processes.
Innovation Solution
The proposed solution involves an MRI system with a first and second group of coils, where the first image is used as training input data and the second image, generated to compensate for sensitivity differences and noise correlations, is used as a label for supervised learning, allowing the post-processing part to receive input and generate a denoised image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used for denoising MRI images, then denoising performance can be improved, but true images or label images with similar noise characteristics are required for training which are difficult or impossible to prepare
Solution Approach 1:
The patent introduces an intermediary approach by using a second group of coils to acquire training data indirectly. Instead of requiring true images or carefully matched label images, the system uses images from a different coil group that serves as a practical surrogate for training, eliminating the need for difficult-to-obtain true images while still enabling supervised learning.
Solution Approach 2:
The patent changes the parameter of coil group selection for training data acquisition. By acquiring training images using a second group of coils different from those used for the input image, the system transforms the training data acquisition process into a more feasible operation that doesn't require true images or precise noise matching.
2Ease of manufacture
If images from different coil groups are used for training, then training data acquisition becomes feasible, but coil sensitivity differences and noise correlations must be compensated
Solution Approach 1:
The patent implements feedback mechanisms by calculating coil sensitivity maps and using them to compensate for sensitivity differences. The system continuously adjusts the training data based on measured sensitivity variations, ensuring that the different coil groups can be effectively used for training despite their inherent sensitivity differences.
Solution Approach 2:
The patent applies counterbalancing techniques to offset the harmful effects of coil sensitivity differences and noise correlations. By introducing compensation algorithms that counteract these differences, the system enables the use of images from different coil groups without being hindered by their sensitivity variations.
3Measurement precision
If true images are used for supervised learning training, then optimal denoising can be achieved, but true images are difficult or impossible to obtain
Solution Approach 1:
The patent creates a practical copy or surrogate for true images by using images acquired from a second group of coils as training labels. This copying approach replaces the unavailable true images with accessible alternative images that can serve the same training function, making supervised learning feasible in practical scenarios.
Data Source
AI summary
Disclosed is a training method including outputting an MRI signal from a plurality of coils included in an MRI scanner and performing, by a computing device, supervised learning on a post-processing part included in the computing device by using, as training input data, a first image generated using a first group of coils among the plurality of coils and using, as a label, a second image generated using a second group of coils among the plurality of coils.