Manifold Learning Image Reconstruction for Undersampled Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image reconstruction techniques in medical imaging modalities like MRI and CT are constrained by the Nyquist criterion, leading to artifacts such as aliasing and streaking, especially in undersampled data, and require pre-designed acquisition protocols to compensate for data shortcomings.
Innovation Solution
A neural network-based system and method that transforms raw sensor data from various imaging domains into image domains without predesigned acquisition constraints, using data-driven manifold learning to inform future data acquisition strategies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image reconstruction techniques (filtered backprojection, Fourier transform) are used, then the reconstruction process is well-established and computationally efficient, but the Nyquist criterion must be satisfied which requires sufficient sampling density and leads to artifacts (aliasing, streaking) in undersampled data
Solution Approach 1:
The patent replaces conventional mathematical transformation methods (filtered backprojection, Fourier transform) with a neural network-based system. The neural network learns the mapping from sensor data to image space through training, substituting the deterministic mathematical operations with a data-driven computational model that can handle undersampled data without requiring strict adherence to the Nyquist criterion
Solution Approach 2:
The patent changes the fundamental parameters of the reconstruction process by using a neural network with learnable weights and biases instead of fixed mathematical operators. The system transforms the reconstruction problem from a deterministic mathematical transformation to a probabilistic learning problem, where the network parameters are optimized during training to minimize reconstruction error
2Manufacturing precision
If pre-designed acquisition protocols are used to compensate for data shortcomings, then the reconstruction can achieve acceptable quality, but the acquisition process becomes constrained and less flexible
Solution Approach 1:
The patent inverts the traditional approach by not designing acquisition protocols to compensate for limitations in reconstruction algorithms. Instead, the neural network is trained to handle various sampling patterns and undersampled data, allowing flexible acquisition protocols without requiring pre-designed compensation strategies. The reconstruction method adapts to the acquisition protocol rather than the protocol being constrained by the reconstruction method
3Manufacturing precision
If sufficient sampling density is used to satisfy the Nyquist criterion, then reconstruction artifacts are minimized, but the data acquisition time and resource requirements increase
Solution Approach 1:
The patent applies partial action by using fewer samples than the Nyquist criterion requires. The neural network is trained to reconstruct high-quality images from undersampled data, effectively performing reconstruction with partial sampling density. This allows the system to achieve acceptable or superior image quality without acquiring the full amount of data that conventional methods would require
Data Source
AI summary
A system may transform sensor data from a sensor domain to an image domain using data-driven manifold learning techniques which may, for example, be implemented using neural networks. The sensor data may be generated by an image sensor, which may be part of an imaging system. Fully connected layers of a neural network in the system may be applied to the sensor data to apply an activation function to the sensor data. The activation function may be a hyperbolic tangent activation function. Convolutional layers may then be applied that convolve the output of the fully connected layers for high level feature extraction. An output layer may be applied to the output of the convolutional layers to deconvolve the output and produce image data in the image domain.