Contrast-Enhanced Radiology in Frequency Space for Artifact-Free Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating artificial radiological images using machine learning face challenges such as errors in co-registration, high computational complexity, and the risk of stitching artefacts, particularly when using large radiological images.
Innovation Solution
The method involves converting radiological images to frequency space for training and prediction using machine learning models, allowing for more efficient processing and reduced artefacts by separating and encoding contrast information independently of fine structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complete radiological images are used for prediction in real space, then the prediction accuracy is improved, but the calculation complexity and computing power requirements increase rapidly
Solution Approach 1:
The patent replaces the mechanical approach of processing images in real space with a transformation to frequency space (k-space). This substitution fundamentally changes the processing domain, allowing predictions to be made in frequency space where calculations are more efficient. The inverse Fourier transform then converts the frequency space prediction back to real space for final image generation, resolving the contradiction between accuracy and computational complexity.
2Quantity of substance
If radiological images are reduced to partial regions and processed separately, then the memory overload is prevented, but stitching artefacts appear at the interfaces
Solution Approach 1:
The patent replaces the spatial processing approach (dividing images into patches in real space) with frequency space processing. In k-space, the entire image can be processed as a unified structure without physical division, eliminating the boundary issues that cause stitching artefacts. The frequency domain naturally handles global image relationships, allowing memory-efficient processing without compromising image quality at interfaces.
3Measurement precision
If co-registration of radiological images is performed to match pixels/voxels precisely, then the accuracy of artificial image generation is improved, but the method becomes sensitive to registration errors and time-consuming
Solution Approach 1:
The patent replaces the pixel/voxel-level co-registration process in real space with frequency space alignment. In k-space, the Fourier transform properties allow for more robust alignment that is less sensitive to small registration errors. The frequency domain representation inherently handles phase shifts and spatial misalignments more gracefully, reducing both the time required for co-registration and the sensitivity to registration inaccuracies.
4Ease of manufacture
If standard hardware is used for calculating artificial radiological images, then the cost is reduced, but the calculation time becomes excessively long
Solution Approach 1:
The patent changes the fundamental parameter of the processing domain from real space to frequency space. This parameter change transforms computationally intensive operations in real space into more efficient operations in frequency space. The Fourier transform-based approach enables standard hardware to process images faster by exploiting the mathematical properties of frequency domain representations, achieving good calculation times without requiring specialized or expensive hardware.
Data Source
AI summary
The present invention relates to the technical field of producing artificial contrast-enhanced radiological images by way of machine learning methods.