GAN-CIRCLE CT Super-Resolution via Residual and Cycle Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CT imaging technologies face limitations in achieving high resolution due to hardware constraints, resulting in inadequate image quality for applications like early tumor characterization and coronary artery analysis, and existing computation-oriented techniques struggle with ill-posed inverse problems and noise artifacts.
Innovation Solution
A CT super-resolution generative adversarial network (GAN) constrained by an identical, residual, and cycle learning ensemble (GAN-CIRCLE) is employed, utilizing a cascade of neural networks with skip connections and loss functions like adversarial, cycle-consistency, and joint sparsifying transform losses to enhance image resolution and reduce noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If hardware constraints (focal spot size, detector element pitch) are used to achieve CT imaging, then the system is simple and reliable, but the image resolution is limited to millimeter order and cannot achieve micrometer order resolution
Solution Approach 1:
The patent replaces hardware-based resolution improvement (mechanical system with smaller focal spots and detector elements) with a computational approach using deep learning algorithms. The neural network processes low-resolution CT images to generate high-resolution images, substituting physical hardware limitations with software-based super-resolution techniques.
Solution Approach 2:
The patent changes the resolution parameter of CT images through computational processing. By applying super-resolution algorithms, the system transforms low-resolution images (millimeter order) into high-resolution images (micrometer order) without physically changing the imaging hardware parameters.
2Measurement precision
If traditional reconstruction algorithms are used to generate CT images, then the process is simple and fast, but the intrinsic resolution is lower than ideal resolution for early tumor characterization and coronary artery analysis
Solution Approach 1:
The patent replaces traditional reconstruction algorithms with deep learning-based super-resolution algorithms. Instead of using conventional iterative or analytical reconstruction methods, the system employs trained neural networks that have learned to enhance resolution, substituting traditional signal processing with intelligent computational models.
Solution Approach 2:
The patent performs preliminary training of neural networks using paired high-resolution and low-resolution CT images before actual reconstruction. The network is pre-trained to learn the mapping between low and high resolution images, so that during actual use, the pre-trained model can quickly enhance resolution without requiring complex real-time computation.
3Device complexity
If existing computation-oriented techniques are used for super-resolution, then hardware requirements are reduced, but the techniques struggle with ill-posed inverse problems and produce noise artifacts
Solution Approach 1:
The patent incorporates feedback mechanisms through the use of multiple loss functions during neural network training, including adversarial loss, cycle consistency loss, and perceptual loss. These feedback mechanisms guide the network to produce more realistic and accurate high-resolution images by continuously comparing generated images with ground truth and adjusting parameters accordingly.
Solution Approach 2:
The patent uses a composite approach by combining multiple neural network architectures and multiple loss functions into a unified super-resolution framework. The system integrates different computational techniques (adversarial training, cycle consistency, perceptual losses) to create a robust solution that overcomes the limitations of individual methods.
Data Source
AI summary
A system for generating a high resolution (HR) computed tomography (CT) image from a low resolution (LR) CT image is described. The system includes a first generative adversarial network (GAN) and a second GAN. The first GAN includes a first generative neural network (G) configured to receive a training LR image dataset and to generate a corresponding estimated HR image dataset, and a first discriminative neural network (DY) configured to compare a training HR image dataset and the estimated HR image dataset. The second GAN includes a second generative neural network (F) configured to receive the training HR image dataset and to generate a corresponding estimated LR image dataset, and a second discriminative neural network (DX) configured to compare the training LR image dataset and the estimated LR image dataset. The system further includes an optimization module configured to determine an optimization function based, at least in part, on at least one of the estimated HR image dataset and/or the estimated LR image dataset. The optimization function contains at least one loss function. The optimization module is further configured to adjust a plurality of neural network parameters associated with at least one of the first GAN and/or the second GAN, to optimize the optimization function.