Invertible Neural Networks for MRI Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging technologies, particularly those using machine learning, struggle with generating high-quality medical images due to issues like blurriness, loss of details, and high noise levels, especially when dealing with under-sampled MRI data.
Innovation Solution
The use of an invertible neural network (INN) to process medical images, specifically designed to learn a mapping function that can transform sub-optimal images into improved versions by mapping them to a latent space representation of a known probability distribution, thereby enhancing image sharpness, resolution, and reducing noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning technologies are used to generate medical images from under-sampled data, then image generation speed and productivity are improved, but image quality deteriorates with blurriness, loss of details, and high noise levels
Solution Approach 1:
The patent inverts the traditional training approach by training the neural network to map from high-quality fully-sampled images to under-sampled images, rather than the conventional approach of mapping from under-sampled to reconstructed images. This inversion allows the network to learn the degradation process, making the reverse reconstruction more accurate and reducing artifacts, blurriness, and noise in the generated images.
Solution Approach 2:
The patent changes the sampling parameter by using fully-sampled images as training inputs instead of under-sampled images. This parameter change in the training data quality enables the network to learn from high-quality references, improving the reconstruction accuracy and reducing the harmful effects of under-sampling while maintaining fast generation speeds.
2Loss of time
If under-sampled MRI data is used to reduce scanning time, then productivity and patient comfort are improved, but measurement precision and image detail are lost
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network on pairs of fully-sampled and under-sampled images before actual reconstruction. This preliminary training phase enables the network to learn the relationship between different sampling densities, so that during actual use, fast reconstruction with under-sampled data can achieve quality comparable to fully-sampled images without the time penalty.
3Device complexity
If traditional neural network structures are used for image reconstruction, then device complexity is kept manageable, but image quality and reliability deteriorate due to blurriness and noise
Solution Approach 1:
The patent inverts the conventional reconstruction approach by training networks to learn the forward degradation process rather than the inverse reconstruction process. This inversion simplifies the training stability and improves reliability, as the network learns from clear reference images how degradation occurs, making the reverse process more reliable and consistent while maintaining manageable complexity.
Data Source
AI summary
Described herein are systems, methods, and instrumentalities associated with using an invertible neural network to complete various medical imaging tasks. Unlike traditional neural networks that may learn to map input data (e.g., a blurry reconstructed MRI image) to ground truth (e.g., a fully-sampled MRI image), the invertible neural network may be trained to learn a mapping from the ground truth to the input data, and may subsequently apply an inverse of the mapping (e.g., at an inference time) to complete a medical imaging task. The medical imaging task may include, for example, MRI image reconstruction (e.g., to increase the sharpness of a reconstructed MRI image), image denoising, image super-resolution, and/or the like.