Self-Supervised ML for MRI Reconstruction from Undersampled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning methods for image reconstruction face challenges when fully-sampled data is not available due to physiological or physical constraints, making it difficult to train machine learning algorithms effectively for inverse problems in scenarios like MRI reconstruction.
Innovation Solution
A computer-implemented method that trains machine learning algorithms using sub-sampled data by dividing it into training and loss function subsets, allowing the algorithms to enforce data consistency and define loss functions separately, enabling reconstruction from undersampled k-space data without requiring fully-sampled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fully-sampled data is used for training machine learning algorithms, then image reconstruction quality is improved, but data acquisition time and physiological constraints are worsened
Solution Approach 1:
The patent segments the sub-sampled data into two distinct subsets: a first subset used for training the neural network and a second subset used for defining the loss function. This segmentation allows the algorithm to learn from one portion of the data while being evaluated against another, enabling effective training without requiring fully-sampled reference data.
2Productivity
If sub-sampled data is used for training, then data acquisition efficiency is improved, but training effectiveness and reconstruction accuracy are worsened
Solution Approach 1:
The patent introduces a self-supervised learning framework that acts as an intermediary mechanism. By using a loss function defined on the second subset of sub-sampled data, the system creates a self-evaluation mechanism that guides the training process without requiring external fully-sampled reference data, thereby maintaining reconstruction accuracy while using efficient sub-sampled training data.
3Measurement precision
If sub-sampled data is divided into training and loss function subsets, then reconstruction performance is improved, but data processing complexity is worsened
Solution Approach 1:
The patent applies segmentation by dividing sub-sampled data into two functional subsets: one for training the neural network parameters and another for defining the loss function. This segmentation enables the algorithm to achieve better reconstruction performance by separating the learning process from the evaluation process, even though it increases data processing steps.
Data Source
AI summary
Self-supervised training of machine learning (“ML”) algorithms for reconstruction in inverse problems are described. These techniques do not require fully sampled training data. As an example, a physics-based ML reconstruction can be trained without requiring fully-sampled training data. In this way, such ML-based reconstruction algorithms can be trained on existing databases of undersampled images or in a scan-specific manner.