Neural Network X-ray CT Image Reconstruction for Artifact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing X-ray computed tomography (CT) systems for baggage screening, such as 2D EDS and stationary CT EDS, face challenges in detecting hidden dangerous objects due to severe streaking artifacts and high operation times, limiting their effectiveness in rapid screening required for airports and harbors.
Innovation Solution
An image processing method using a neural network learned in both image and sinogram domains, incorporating a convolutional framelet-based neural network with multi-resolution capabilities and bypass connections, to reconstruct high-quality multi-directional X-ray CT images from 9-view CT data, addressing artifact corruption and improving reconstruction speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional filtered-backprojection (FBP) algorithm is used with 9 projection views, then reconstruction speed is fast, but severe streaking artifacts occur degrading image quality
Solution Approach 1:
A neural network is introduced as an intermediary component between the FBP algorithm and the final image output. The neural network processes the initial reconstruction results to remove streaking artifacts while preserving image features, effectively mediating between the fast but artifact-prone FBP method and the requirement for high image quality.
Solution Approach 2:
The patent replaces the traditional iterative mechanical reconstruction process (MBIR) with a neural network-based approach. This substitution maintains fast reconstruction speed while achieving artifact removal through learned patterns rather than iterative computational mechanics.
2Manufacturing precision
If model-based iterative reconstruction (MBIR) is used, then image quality improves by reducing artifacts, but operation time increases making it unsuitable for rapid screening
Solution Approach 1:
The patent substitutes the computationally intensive iterative MBIR process with a pre-trained neural network that performs reconstruction in a single pass. The neural network has learned artifact removal patterns during training, enabling it to achieve MBIR-quality results without the iterative computational overhead.
Solution Approach 2:
The neural network is pre-trained offline using MBIR-generated data, performing the complex iterative reconstruction work in advance. During actual operation, the pre-learned knowledge is applied directly to new data, achieving fast reconstruction without repeating the iterative process.
3Device complexity
If stationary CT EDS with fixed sources and detectors is used, then mechanical complexity is reduced for routine inspection, but detection capability degrades due to limited projection views
Solution Approach 1:
The neural network acts as an intermediary that compensates for the limited 9-view projection data. It learns to reconstruct high-quality images from the sparse views by filling in missing information and removing artifacts, effectively mediating between the simplified stationary hardware and the requirement for reliable detection.
Solution Approach 2:
The patent changes the reconstruction parameter space by introducing learned regularization through the neural network. Instead of relying solely on the physical geometry of multiple projection views, the system uses learned parameters from training data to achieve reliable reconstruction from limited views.
Data Source
AI summary
A method for processing a multi-directional X-ray computed tomography (CT) image using a neural network and an apparatus therefor are provided. The method includes receiving a predetermined number of multi-directional X-ray CT data and reconstructing an image for the multi-directional X-ray CT data using a neural network learned in each of an image domain and a sinogram domain.