Texture Matching GAN for CT Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image enhancement techniques for CT images using generative adversarial networks (GANs) fail to provide quantitative evaluation of texture and often introduce inaccurate or unreal image details, known as hallucinations, which are undesirable in medical imaging.
Innovation Solution
The proposed system employs a Texture Matching Generative Adversarial Network (TMGAN) that separates texture from anatomical structures by processing two independent noisy samples of images through a Siamese network, allowing the discriminator to score and match the texture to a target distribution, thereby producing denoised or sharpened images with desired textures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing GAN approaches are used to enhance CT images, then image denoising or sharpening is improved, but inaccurate or unreal image details (hallucinations) are introduced
Solution Approach 1:
The patent segments the image enhancement task by separating texture processing from anatomical structure processing. The generator network processes only texture information independently, while anatomical structures are preserved from the original image, preventing hallucinations in anatomical regions.
Solution Approach 2:
The patent introduces a texture extractor as an intermediary component that isolates texture information from the input image before feeding it to the generator. This intermediary step ensures that only texture details are synthesized, preventing the introduction of false anatomical structures.
2Manufacturing precision
If existing GAN approaches are used to enhance CT images, then image texture is improved, but quantitative evaluation of texture is not provided
Solution Approach 1:
The patent implements a feedback mechanism where the discriminator evaluates the generated texture and provides quantitative feedback to the generator. This closed-loop system enables continuous optimization of texture quality with measurable performance metrics.
Solution Approach 2:
The patent transforms the texture evaluation problem into a parameter-based assessment by using the discriminator to score texture samples against a target distribution, converting qualitative texture assessment into quantitative parameter comparison.
3Ease of operation
If existing GAN approaches are used to enhance CT images, then image processing is simplified, but texture matching to target distribution is not achieved
Solution Approach 1:
The patent performs preliminary action by pre-processing the input image to extract texture components before the main generation process. This preliminary separation simplifies the subsequent generation task while enabling precise texture matching to the target distribution.
Data Source
AI summary
One or more systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to image enhancement using a generative adversarial network (GAN). The computer-implemented system can comprise a memory that can store computer-executable components. The computer-implemented system can further comprise a processor that can execute the computer-executable components stored in the memory, wherein the computer-executable components can comprise a training component that can train a discriminator of the GAN to score a texture of a CT image, wherein the texture can be derived from a difference of two conditionally independent estimates produced by a generator of the GAN by respectively processing two independent noisy samples of images.