Self-Supervised MRI Reconstruction With Weighted Loss Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-resolution magnetic resonance imaging (MRI) acquisition with large coverage is time-consuming, leading to reduced patient compliance and clinical workflow inefficiency, while fast MRI methods result in image reconstruction artifacts and high noise, and existing deep learning techniques require groundtruth data for training.
Innovation Solution
A self-supervised deep learning model (SelfDEQ) for MRI reconstruction that uses a model-based implicit neural network trained on undersampled and noisy measurements, incorporating a weighted loss function and Jacobian-free backward pass to optimize training without groundtruth data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If fast MRI is performed by sparsely sampling magnetic resonance k-space below the Nyquist rate, then acquisition time is reduced, but image reconstruction artifacts and high noise occur
Solution Approach 1:
The patent introduces a self-supervised deep learning model as an intermediary between the undersampled k-space data and the final reconstructed image. This model learns the mapping from undersampled to fully-sampled k-space representations, effectively mediating the reconstruction process to reduce artifacts and noise while maintaining acceleration benefits
Solution Approach 2:
The patent transforms the reconstruction problem by changing the parameter space - instead of directly reconstructing from undersampled data, the method learns optimal k-space sampling patterns and reconstruction weights as trainable parameters, allowing the system to adaptively optimize the trade-off between sampling rate and image quality
2Manufacturing precision
If compressed sensing reconstruction is used to reconstruct undersampled MR data, then image quality is improved, but computational time consumption increases
Solution Approach 1:
The patent replaces the iterative mechanical optimization process of compressed sensing with a learned direct mapping through deep neural networks. The self-supervised model learns the reconstruction transformation during training, enabling rapid inference without requiring iterative optimization during actual reconstruction, thus substituting computational iteration with pre-learned transformation
3Manufacturing precision
If deep learning methods are used for MRI reconstruction, then reconstruction quality is improved, but groundtruth data is required for training
Solution Approach 1:
The patent implements self-service learning by formulating a self-supervised training objective where the model learns to reconstruct images from undersampled data without requiring groundtruth labels. The training process uses the undersampled data itself to supervise learning through consistency constraints and learned priors, making the system self-sufficient and eliminating the need for expensive fully-sampled training datasets
Data Source
AI summary
A system for image reconstruction includes an input for receiving image data, a processor, and a memory. The memory stores instructions that cause the processor to reconstruct an image from the image data using a self-supervised deep learning model.