GAN-Based CT Reconstruction with Gradient Guidance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CT super-resolution methods rely on simulated or phantom image pairs, leading to poor performance in real clinical settings due to misalignment and require increased scanning time and radiation dose, which are not suitable for direct application in clinical environments.
Innovation Solution
A GAN-based framework with gradient guidance that uses clinical data pairs and incorporates a gradient guidance branch to directly learn the mapping for ultra-high resolution CT image reconstruction, employing a hybrid loss function combining pixel-wise, perceptual, and adversarial losses to generate reliable anatomical structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If hardware-oriented methods are used to improve spatial resolution, then manufacturing precision is improved, but loss of time increases and object-affected harmful factors increase
Solution Approach 1:
The patent replaces hardware-oriented mechanical resolution improvement methods with a deep learning-based image processing system. The neural network model processes existing CT images to enhance spatial resolution without requiring additional scanning time or hardware modifications, thus resolving the contradiction between manufacturing precision and loss of time.
Solution Approach 2:
The patent changes the parameter of image resolution through software-based deep learning processing rather than hardware adjustments. The neural network model transforms low-resolution CT images into high-resolution images by learning from training data, achieving resolution improvement without increasing scanning time.
2Manufacturing precision
If hardware-oriented methods are used to improve spatial resolution, then manufacturing precision is improved, but object-affected harmful factors increase
Solution Approach 1:
The patent substitutes hardware-based resolution enhancement that requires increased radiation dose with a software-based deep learning approach. The neural network model enhances image quality by processing existing images, eliminating the need for additional radiation exposure while maintaining improved spatial resolution.
3Manufacturing precision
If deep learning based super-resolution methods are used, then manufacturing precision is improved, but reliability deteriorates due to poor performance in real clinical settings
Solution Approach 1:
The patent applies preliminary action by extensively training the neural network model on large datasets of paired low-resolution and high-resolution CT images before deployment. This pre-training on diverse clinical data ensures the model generalizes well to real-world scenarios, improving reliability while maintaining high resolution enhancement capability.
Solution Approach 2:
The patent incorporates feedback mechanisms through loss functions that compare generated high-resolution images with ground truth images during training. This feedback loop continuously optimizes the model to produce clinically accurate results, ensuring reliability in real-world applications.
4Manufacturing precision
If conventional super-resolution methods are used, then manufacturing precision is improved, but loss of information increases due to fake details
Solution Approach 1:
The patent uses perceptual loss and adversarial loss as feedback mechanisms during training to ensure generated high-frequency details are anatomically accurate rather than fabricated. The adversarial network component specifically targets the generation of realistic textures and structures, reducing information loss by preventing fake detail generation.
Data Source
AI summary
A computer-implemented method is provided for ultra-high resolution computed tomography. The method comprises: acquiring, using computed tomography (CT), a medical image of a subject, the medical image has a lower resolution; and processing the medical image, with aid of a deep learning network model, to reconstruct an ultra-high resolution medical image, where the deep learning network model is trained using a generative adversarial network (GAN)-based framework with a gradient guidance.