Complex Denoising CNN for Low-Field MRI Signal Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current denoising algorithms for MRI images, such as non-local means and block-matching and 3D filtering, struggle with small lesion regions and non-uniform noise patterns, especially in multi-coil MRI data, due to their reliance on repetitive structures and sensitivity to noise distribution changes, which limits their effectiveness in clinical applications.
Innovation Solution
A complex de-noising convolutional neural network (C-DnCNN) is employed, which acquires complex MRI data and iteratively updates parameter settings to predict residual images from noisy input images, allowing for the removal of latent clean images without requiring high-quality training data, thus addressing the limitations of traditional methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional filter-based denoising methods (NLM, BM3D) are used, then computational simplicity is maintained, but denoising performance deteriorates on small lesion regions and non-uniform noise patterns
Solution Approach 1:
The patent replaces traditional mechanical filter-based denoising systems (NLM, BM3D) with a deep learning-based neural network system. The C-DnCNN uses learned complex-valued filters and iterative refinement to achieve superior denoising performance on challenging MRI data with non-uniform noise and small lesions, while maintaining computational efficiency through the neural network architecture.
Solution Approach 2:
The patent introduces complex-valued parameters and iterative refinement parameters to enhance the denoising capability. The C-DnCNN processes complex-valued MRI data and uses iterative updates to progressively improve denoising performance, allowing the system to adapt to non-uniform noise patterns and preserve fine structural details that traditional real-valued filters cannot handle effectively.
2Productivity
If aggressive acceleration strategies are used, then scan time is reduced, but signal-to-noise ratio deteriorates
Solution Approach 1:
The patent converts the harmful effect of noise introduced by aggressive acceleration strategies into a beneficial outcome. The C-DnCNN is specifically designed to handle the non-uniform noise patterns that result from parallel imaging and B1 inhomogeneity, transforming the previously problematic noise into manageable statistical characteristics that the neural network can effectively model and remove, thereby enabling faster scans without sacrificing image quality.
Solution Approach 2:
The patent introduces the C-DnCNN as an intermediary processing step between data acquisition and final image reconstruction. This neural network mediator handles the noise compensation task, allowing the system to use aggressive acceleration strategies during acquisition while the C-DnCNN subsequently removes the introduced noise, effectively decoupling the scan speed optimization from the image quality degradation.
3Reliability
If more averages and lower bandwidth are used, then signal-to-noise ratio is improved, but scan time increases
Solution Approach 1:
The patent uses a learned model (C-DnCNN) that has been trained to replicate the denoising效果 of traditional methods while being computationally more efficient. The neural network learns from training data to produce high-quality denoised images without requiring multiple averages or prolonged scan times, effectively copying the beneficial outcome of traditional denoising approaches while avoiding their time-consuming requirements.
4Reliability
If complex-valued operations are used, then denoising performance on non-uniform noise is improved, but computational complexity increases
Solution Approach 1:
The patent segments the complex-valued processing into distinct manageable components within the neural network architecture. The C-DnCNN separates complex multiplication, convolution, and iterative refinement into discrete operational stages, making the computational complexity tractable while preserving the benefits of complex-valued operations for handling non-uniform noise patterns in MRI data.
Data Source
AI summary
MR image data can be improved by using a complex de-noising convolutional neural network such as a non-blind C-DnCNN, a network for MRI denoising that leverages complex-valued data with phase information and noise level information to improve denoising performance in various settings. The proposed method achieved superior performance on both simulated and in vivo testing data compared to other algorithms. The utilization of complex-valued operations allows the network to better exploit the complex-valued MRI data and preserve the phase information. The MR image data is subject to complex de-noising operations directly and simultaneously on both real and imaginary parts of the image data. Complex and real values are also utilized for block normalization and rectified linear units applied to the noisy image data. A residual image is predicted by the C-DnCNN and a clean MR image is available for extraction.