Neural Network Patch Reconstruction for MRI Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies face challenges in reconstructing high-quality images due to under-sampling, which results in artifacts like aliasing and streaking, especially in modalities such as MRI and CT, where the Nyquist criterion is often violated, leading to suboptimal image quality and increased scan times.
Innovation Solution
The use of data-driven manifold learning techniques, specifically the AUTOMAP framework, allows for the transformation of raw sensor data into medical images without preconceived constraints on data acquisition, enabling the reconstruction of images even when the Nyquist criterion is not met, and can inform future data acquisition strategies to improve image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional imaging methods are used to reconstruct images from sensor data, then the imaging process follows established protocols, but the images suffer from artifacts like aliasing and streaking due to under-sampling and violation of the Nyquist criterion
Solution Approach 1:
The patent replaces traditional mechanical image reconstruction methods with a neural network-based system. The neural network learns the complex mapping between sensor data and images during training, enabling it to reconstruct high-quality images from under-sampled data without the artifacts that plague traditional methods. This substitution of the reconstruction mechanism resolves the contradiction by eliminating artifacts while maintaining image quality.
Solution Approach 2:
The patent changes the approach to image reconstruction by transforming the problem from a deterministic mathematical inversion to a data-driven learning problem. By training neural networks on pairs of sensor data and corresponding images, the system learns optimal reconstruction parameters and mappings, enabling high-quality image generation even when the Nyquist criterion is violated and traditional reconstruction fails.
2Productivity
If the Nyquist criterion is violated to reduce scan time, then image capture is accelerated, but image quality deteriorates due to under-sampling errors
Solution Approach 1:
The neural network incorporates feedback mechanisms during training, where it learns from the relationship between sensor data and corresponding images. This feedback loop enables the network to compensate for under-sampling errors and reconstruct high-quality images even when scan time is reduced. The feedback during training allows the system to internalize the consequences of under-sampling and develop reconstruction strategies that maintain image quality at accelerated scan speeds.
Solution Approach 2:
The patent uses copying by training neural networks on pairs of sensor data and corresponding high-quality images. The network learns to copy the essential features and relationships from the training data, enabling it to generate accurate reconstructions from under-sampled input data. This copying process allows the system to produce high-quality images even when the input data is incomplete due to reduced scanning.
3Manufacturing precision
If data-driven manifold learning techniques are applied to transform sensor data into images, then image quality improves and artifacts are reduced, but the system complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the image reconstruction task into manageable components through patch-based processing. Instead of processing entire images at once, the system processes smaller patches of sensor data and corresponding image regions through the neural network. This segmentation reduces the computational complexity of each individual transformation step while maintaining overall image quality, making the data-driven approach more feasible despite the inherent complexity of manifold learning.
4Productivity
If patch-based processing is used to handle complex input datasets, then computational efficiency improves, but the ability to capture global image context may be compromised
Solution Approach 1:
The patent applies dimensionality change by operating on image patches in multiple dimensions - both spatially (processing different patch locations) and hierarchically (processing patches at different scales or resolutions). The neural network processes patches through multiple layers and combinations, effectively reconstructing global image context from local patch information. This multi-dimensional processing approach maintains computational efficiency while preserving global context through the hierarchical assembly of patch-level transformations.
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 enables the generation of high-quality medical images with reduced noise and artifacts, improved noise immunity, and accelerated image capture, while being flexible enough to compensate for hardware imperfections and under-sampling errors, thus enhancing patient throughput and comfort.
Implementation Method 1
applying a Fourier transform to the first image patch to transform the first image patch into a first sensor data patch
Data Source
AI summary
Systems, methods, and media for patch-based medical image generation for complex input datasets. Patch-based medical image generation can include creating a training dataset with an image patch and corresponding sensor data patch and training a neural network using the training dataset. Then, sensor data acquired from a patient using a medical imaging system can be applied as input to the neural network, and a medical image of the patient can be generated based on an output of the neural network.