Pretrained 3D Neural Networks for MR Super-Resolution and Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing magnetic resonance (MR) techniques face challenges in extracting accurate 3D properties from noisy and low-resolution 2D slices, leading to time-consuming and non-quantitative measurements.
Innovation Solution
A computer system uses a pretrained neural network, such as a denoising diffusion probabilistic model (DDPM) or a 3D generative adversarial network (GAN), to enhance MR measurements by increasing resolution and reducing noise, leveraging techniques like convolutional neural networks (CNNs) and transformers for parameter computation and image reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional MR techniques are used to extract 3D properties from 2D slices, then measurement accuracy is maintained, but measurement time increases and productivity decreases
Solution Approach 1:
The neural network is pre-trained on a large dataset of MR images with corresponding 3D properties before deployment. This preliminary training phase enables the network to learn complex mapping relationships from 2D slices to 3D properties, so that during actual measurement, the pre-learned knowledge can be rapidly applied without time-consuming iterative processing
Solution Approach 2:
The patent replaces traditional mechanical/iterative image processing methods with a neural network-based computational approach. The neural network directly maps 2D slice data to 3D property predictions, substituting complex iterative reconstruction algorithms with a single-pass inference process that maintains accuracy while dramatically reducing computation time
2Reliability
If traditional MR techniques are used to process noisy 2D slices, then noise reduction is achieved, but processing time increases
Solution Approach 1:
The neural network is pre-trained on noisy MR images with ground truth labels, learning optimal denoising patterns during the training phase. This allows the network to perform noise reduction in a single forward pass during inference, eliminating the need for time-consuming iterative denoising algorithms while maintaining high signal-to-noise ratio
Solution Approach 2:
The neural network creates a cleaned version of the noisy input image by learning the mapping from noisy to clean images during training. This copied clean image can be generated directly from the noisy input without iterative processing, providing both noise reduction and time efficiency
3Manufacturing precision
If conventional MR imaging is used to achieve adequate resolution, then image quality is maintained, but scan time increases
Solution Approach 1:
The patent transitions from 2D slice-based processing to 3D volumetric processing by using 3D convolutional neural networks or 3D U-Net architectures. This dimensional transformation allows the network to capture spatial correlations across multiple slices simultaneously, achieving high-resolution 3D reconstruction from fewer 2D acquisitions, thereby reducing scan time while maintaining or improving image quality
Solution Approach 2:
The neural network generates high-resolution 3D image copies from lower-resolution input data. By learning the mapping from low-resolution to high-resolution images during training, the network can synthesize detailed 3D structures without requiring high-resolution acquisitions, thus reducing scan time while maintaining image quality
Data Source
AI summary
A computer system that increases resolution and/or reduces noise (and/or artifacts) in measurements is described. During operation, the computer system may obtain information specifying simulated MR measurements and at least an MR measurement. Then, the computer system may increase the resolution and/or reduces the noise in at least the MR measurement or the simulated MR measurements using a pretrained neural network, where the pretrained neural network includes a three-dimensional (3D) generative neural network (GAN) in series with a patch discriminator. Note that the patch discriminator may output information indicating whether an input to the pretrained neural network is a real MR measurement or a simulated MR measurement. Moreover, the simulated MR measurements may have a lower resolution than at least the MR measurement.