MRI Denoising Training Data via Non-Linear Noise Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based image denoising methods for magnetic resonance imaging (MRI) face challenges in generating training datasets that accurately represent real-world noise distributions, leading to less accurate denoising results due to simplified noise models that do not account for variations in noise across different contrasts and spatial locations within MRI images.
Innovation Solution
A method and system that generate noisy training images by splitting original images into noise-only and image components, using MR reconstruction steps to produce images with realistic noise characteristics, and combining zero-mean Gaussian noise images in a non-linear manner to create training images that reflect natural MR noise distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If uniform Gaussian noise or uniform Rician noise is added to target images using simple noise models, then the training dataset generation is simplified and faster, but the noise distribution does not accurately represent real-world MR noise characteristics, leading to reduced denoising accuracy
Solution Approach 1:
The patent applies local quality by implementing region-specific noise modeling that differentiates between tissue regions (high signal) and background regions (low signal). Different noise distributions are applied to different spatial locations within the image, with Gaussian noise for tissue regions and Rician noise for background regions, thereby accurately representing the non-uniform noise characteristics of real MR images while maintaining training effectiveness
Solution Approach 2:
The patent changes noise parameters by dynamically adjusting noise standard deviation based on local signal intensity and region type. The noise model transitions between Gaussian and Rician distributions based on signal-to-noise ratio thresholds, allowing the training data to reflect the actual parameter variations encountered in clinical MR imaging across different contrasts and anatomical regions
2Measurement precision
If multiple region-specific noise techniques with accurate noise distributions are implemented, then the denoising accuracy is improved, but the training dataset generation process becomes more complex and computationally intensive
Solution Approach 1:
The patent segments the image into different regions (tissue vs. background) based on signal intensity thresholds and applies appropriate noise models to each segment. This segmentation approach enables accurate region-specific noise characterization while organizing the complex processing into manageable steps: region identification, noise parameter selection, and localized noise application
Solution Approach 2:
The patent performs preliminary actions by pre-calculating noise parameters and establishing region classification criteria before generating training pairs. Noise standard deviations and distribution types are determined in advance based on image contrast and region type, allowing the actual noise addition process to proceed efficiently without real-time complex computations
Data Source
AI summary
A synthetically generated noise image is generated from at least one high signal-to-noise ratio target image and at least two zero-mean Gaussian noise images, scaled according to a noise scale. The images are combined in a non-linear manner to produce the synthetically generated noise image which can be used as a training image in a machine learning-based system that “denoises” images. The process can be repeated for a number of different noise scales to produce a set of training images. In one embodiment, the synthetically generated noise image IN is generated according to:IN=√{square root over (I12+I22)}where I is the original target image, I1=I+pG1 and I2=pG2, and where G1 and G2 are zero-mean Gaussian noise images, and p is the noise scale.