Progressive GAN for MRI Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current MRI reconstruction methods using compressed sensing and deep learning face challenges in generating high-resolution images due to limited training data and difficulty in generalizing to different noise levels, and unsupervised learning techniques like GANs struggle with training for high-resolution image generation.
Innovation Solution
A progressive generative adversarial network (PGAN) is employed for MRI reconstruction, where the generator and discriminator are progressively trained starting from low resolution, adding layers to model finer details, enabling better handling of high-resolution information and image denoising, and used as a regularizer in iterative reconstruction processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If compressed sensing reconstruction is used to accelerate MRI scanning, then scanning speed is improved, but image quality deteriorates due to unfolding artifacts
Solution Approach 1:
The patent transforms the traditional pixel-wise loss function into a perceptual loss function that operates in feature space rather than pixel space. This parameter change in the optimization objective allows the model to prioritize perceptually important features over exact pixel matching, thereby reducing artifacts while maintaining scanning acceleration
Solution Approach 2:
The patent introduces a pre-trained deep network as an intermediary between the compressed sensing reconstruction and the final image output. This intermediary network processes the reconstructed image to enhance perceptual quality, acting as a mediator that transforms mathematically accurate but visually degraded images into perceptually superior images
2Manufacturing precision
If supervised deep learning is used for reconstruction, then reconstruction quality is improved, but training data requirements increase making it difficult and costly to build large datasets
Solution Approach 1:
The patent employs unsupervised learning where the model learns to reconstruct images without requiring paired ground truth images. The network serves itself by learning from the statistical properties of the data and the constraints of the forward model, eliminating the need for extensive manually annotated training datasets while maintaining reconstruction quality
Solution Approach 2:
Instead of using the traditional supervised approach where ground truth images guide the learning process, the patent inverts the paradigm by using unsupervised learning with perceptual loss that compares feature representations rather than pixel values. This inversion allows learning without direct access to ground truth, reducing data requirements
3Adaptability or versatility
If GAN is used for unsupervised learning, then ground truth is not required, but training for high-resolution image generation becomes difficult
Solution Approach 1:
The patent segments the training process into distinct stages: first training on low-resolution images to learn coarse structures, then progressively training on higher-resolution images to learn fine details. This segmentation of the training process by resolution level makes high-resolution GAN training feasible by breaking it into manageable steps
4Manufacturing precision
If progressive GAN is used for high-resolution generation, then image quality is improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary training on low-resolution images before proceeding to high-resolution training. This preliminary action allows the model to learn fundamental reconstruction patterns and features at lower computational cost, preparing the network for subsequent high-resolution training and reducing overall training time
Data Source
AI summary
For reconstruction in medical imaging, such as reconstruction in MR imaging, a high-resolution image is reconstructed using a generator of a progressive generative adversarial network (PGAN or progressive GAN). In machine training the network, both the generator and discriminator of the GAN are grown progressively: starting from a low resolution, new layers are added that model finer details as training progresses. The resulting generator may be better able to handle high-resolution information than a generator of a GAN.