Trained Generator Function for X-ray Image Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In medical X-ray imaging, there is a challenge in balancing image quality and X-ray dose, as noise suppression methods can alter image appearance and introduce artifacts, and optimizing signal-to-noise ratio through protocol selection can change pixel values, making it difficult for further processing by trained algorithms, especially when data is limited.
Innovation Solution
A computer-implemented method and system that uses a trained generator function to reduce noise in X-ray image datasets by applying a parameter-based transformation, adjusting the generator function based on comparisons between training and resultant image datasets, to produce images with lower noise levels while maintaining accurate representation of the volume of interest.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If noise suppression methods are used to increase image quality, then image quality is improved, but image appearance is changed and artifacts are introduced
Solution Approach 1:
The generator function is trained in advance on a large dataset of X-ray images with various noise levels and protocols. This preliminary training enables the function to learn optimal noise reduction patterns without altering image appearance, which can then be applied to new images without requiring additional post-processing that might introduce artifacts.
Solution Approach 2:
The generator function acts as an intermediary between the noisy input image and the desired clean output image. Instead of directly applying noise suppression filters that alter appearance, the generator function transforms the noisy image through learned mappings that preserve anatomical structures while removing noise, thus maintaining image fidelity.
2Object-affected harmful factors
If X-ray dose is reduced, then patient radiation exposure is minimized, but image quality deteriorates and noise level increases
Solution Approach 1:
The generator function converts the harmful effect of noise in low-dose images into a benefit by learning to recognize and remove noise patterns while preserving diagnostic information. The function is trained specifically on low-dose images with high noise levels, enabling it to transform these noisy images into quality images that rival or exceed high-dose image quality.
Solution Approach 2:
The generator function learns to map images across different noise levels and X-ray protocols by changing the effective noise parameter. When applied to low-dose images, it transforms them to have the quality characteristics of high-dose images, effectively changing the noise parameter from high to low while maintaining the same anatomical information.
3Measurement precision
If protocols are optimized to improve signal-to-noise ratio, then image quality is improved, but pixel values change making further processing difficult
Solution Approach 1:
The generator function is designed to be protocol-agnostic and can process images from different X-ray protocols and apparatuses. By training on diverse datasets encompassing multiple protocols, the function learns universal features that are consistent across different imaging conditions, making the output compatible with various trained algorithms regardless of the input protocol.
Solution Approach 2:
The generator function performs preliminary normalization and standardization of pixel values during the noise reduction process. By learning the relationships between different protocols during training, it outputs images with standardized pixel value distributions that are compatible with downstream algorithms, eliminating the need for protocol-specific processing adjustments.
Data Source
AI summary
A method is for providing a resultant image dataset for a volume of interest. In an embodiment, an X-ray image dataset for the volume of interest is received, having a first noise level. A trained generator function is received, a parameter of which is based on a first and a second training image dataset for a training volume of interest. The first and the second training image dataset include a first training noise level. In addition, a resultant image dataset for the volume of interest is determined by applying the trained generator function to input data comprising the X-ray image dataset, the resultant image dataset has a second noise level less than the first noise level. In addition, the resultant image dataset is provided. As such, a higher noise level can be accepted and/or a lower X-ray dose can be used in the acquisition of the X-ray image dataset.