CT Artifact Correction via Neural Network Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computed tomography (CT) imaging techniques face challenges in correcting artifacts such as windmill and cone beam artifacts, which result in high-frequency portions along the system axis and low-frequency portions in the image plane, leading to image distortion and loss of resolution.
Innovation Solution
An iterative reconstruction method using a trained function, specifically a neural network-based approach, is applied to correct three-dimensional volume image data, allowing for the identification and removal of artifacts without significant hardware upgrades, thereby improving image quality and resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If non-linear filters are used to remove artifacts from volume image data, then artifact correction is achieved, but other structures with similar signatures are also filtered causing image information loss
Solution Approach 1:
The trained function is applied locally to specific regions of the volume image data where artifacts are detected. The system identifies artifact-affected regions and applies correction only to those areas, preserving the original image information in non-affected regions. This localized approach prevents the loss of healthy tissue structures that would occur with global filtering.
Solution Approach 2:
The system uses a trained function that has been pre-trained on artifact-free reference data to recognize and correct artifacts. The correction process incorporates feedback by comparing the corrected regions with the trained model's expectations, allowing iterative refinement that distinguishes between artifacts and actual anatomical structures, thereby preventing information loss.
2Reliability
If thicker slices are reconstructed to fulfill the sampling theorem, then sampling requirements are met, but image resolution in the system axis direction is reduced
Solution Approach 1:
The system performs preliminary correction of sub-sampling artifacts using the trained function before final image reconstruction. By addressing the sampling issues in the preprocessing stage through learned corrections rather than traditional thick-slicing, the system maintains fine resolution while ensuring sampling theorem compliance.
Solution Approach 2:
The patent replaces the mechanical approach of physical thick-slicing (combining multiple detector rows) with a computational approach using a trained neural network function. This substitution allows the system to fulfill sampling requirements through software-based artifact correction rather than hardware-based slice thickening, thereby preserving image resolution.
3Reliability
If a spring focus CT device is used to approximately fulfill the sampling theorem, then sampling requirements are improved, but device structural complexity and cost increase
Solution Approach 1:
The system replaces the complex mechanical spring focus mechanism with a computational solution using a trained function. The trained model learns to correct sub-sampling artifacts from standard CT hardware, eliminating the need for expensive and complex spring focus mechanisms while achieving comparable or superior artifact correction results.
Solution Approach 2:
Instead of modifying the physical CT device with spring focus mechanisms, the system creates a virtual correction layer through a trained function that copies and processes the raw data. This computational copy allows artifact correction without altering the underlying hardware, thereby avoiding increased device complexity and cost.
4Object-affected harmful factors
If traditional filter-based methods are used for artifact correction, then artifacts are removed, but computational effort and processing time increase
Solution Approach 1:
The system performs preliminary training of the function offline using artifact-free reference data. Once trained, the function can be rapidly applied to correct artifacts in clinical scans without requiring intensive computational resources during actual operation. This pre-computation approach significantly reduces processing time compared to traditional iterative filter-based methods.
Solution Approach 2:
The patent transforms the artifact correction problem from a complex iterative filtering process with many adjustable parameters into a direct application of a trained function with fixed parameters. By changing from a multi-parameter optimization problem to a single-function application, the system achieves faster processing while maintaining correction effectiveness.
Data Source
AI summary
A method for the artifact correction of three-dimensional volume image data of an object is disclosed. In an embodiment, the method includes receiving first volume image data via a first interface, the first volume image data being based on projection measurement data acquired via a computed tomography device, the computed tomography device including a system axis, and the first volume image data including an artifact including high-frequency first portions in a direction of a system axis and including second portions, being low-frequency relative to the high-frequency first portions, in a plane perpendicular to the system axis; ascertaining, via a computing unit, artifact-corrected second volume image data by applying a trained function to the first volume image data received; and outputting the artifact-corrected second volume image data via a second interface.