Machine Learning Quantitative MRI Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current quantitative MRI techniques face challenges in accurately measuring tissue parameters like T1 and T2 relaxation times due to oversimplification of signal models, which can lead to systematic errors and increased numerical complexity, making it difficult to obtain precise and robust quantitative maps.
Innovation Solution
A machine learning method and system that learns the relationship between signal intensities and gold-standard values, allowing for model-free characterization of biological tissues and generation of parametric maps, incorporating unwanted effects that traditional models cannot handle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a simple signal model is used for quantitative mapping, then the model fitting is robust and computationally efficient, but the measurement precision deteriorates due to systematic errors from omitted physical effects
Solution Approach 1:
A deep learning network is introduced as an intermediary between the simple acquisition sequence and the quantitative parameter estimation. The network learns the complex mapping from signal data to tissue parameters during training using reference data, then applies this learned mapping during inference. This intermediary captures complex signal behaviors that simple analytical models miss, while keeping the actual measurement sequence simple and fast.
Solution Approach 2:
The deep learning network is trained in advance using a training dataset containing reference values obtained from gold-standard sequences. This preliminary training phase allows the network to learn the relationship between signal data and tissue parameters, including complex effects like stimulated echoes and magnetization transfer, before being deployed for actual quantitative mapping.
2Measurement precision
If a complex signal model is used to accurately describe magnetization behavior, then the measurement precision improves, but the device complexity and computational requirements increase significantly
Solution Approach 1:
The complex analytical signal model is replaced with a data-driven deep learning model. Instead of using complex mathematical equations to describe magnetization behavior, the system uses a neural network that has learned the mapping from signal data to tissue parameters from training data. This substitution maintains accuracy while simplifying the actual measurement and computation during inference.
Solution Approach 2:
The complex relationship between signal data and tissue parameters, which would require complex analytical models to describe, is captured and stored as learned weights and biases in the deep learning network during training. This copied knowledge allows the system to accurately estimate parameters without needing to explicitly model complex physical effects during the actual measurement.
3Measurement precision
If a complex signal model with many independent variables is used, then the measurement precision improves, but the difficulty of detecting and measuring increases due to the ill-posed nature of the fitting problem
Solution Approach 1:
The deep learning network is trained in advance using a training dataset containing reference values obtained from gold-standard sequences. This preliminary training phase allows the network to learn the relationship between signal data and tissue parameters, including complex effects like stimulated echoes and magnetization transfer, before being deployed for actual quantitative mapping.
Solution Approach 2:
A deep learning network is introduced as an intermediary between the simple acquisition sequence and the quantitative parameter estimation. The network learns the complex mapping from signal data to tissue parameters during training using reference data, then applies this learned mapping during inference. This intermediary captures complex signal behaviors that simple analytical models miss, while keeping the actual measurement sequence simple and fast.
4Measurement precision
If gold-standard sequences are used to obtain reference values for model validation, then the measurement precision is improved, but the loss of time increases due to the long acquisition duration
Solution Approach 1:
The deep learning network is trained in advance using a training dataset containing reference values obtained from gold-standard sequences. This preliminary training phase allows the network to learn the relationship between signal data and tissue parameters, including complex effects like stimulated echoes and magnetization transfer, before being deployed for actual quantitative mapping.
Solution Approach 2:
The overall process is segmented into two phases: (1) an offline training phase where gold-standard sequences are used to create reference data for training the deep learning network, and (2) an online inference phase where the trained network rapidly estimates parameters from simple acquisition sequences. This segmentation allows accurate reference data to be used only when necessary for training, while fast estimation is used for actual measurements.
Data Source
AI summary
A system and a method determine a value for a parameter. Reference values for the parameter are determined from a group of objects. A first technique is used by the system for determining for each object the reference value from a first set of data. A learning dataset is created by associating for each object of the group of objects a second set of data and the reference value. The second set of data is acquired by the system according to a second technique for determining values of the parameter and is configured for enabling a determination of the parameter. A machine learning technique trained on the learning dataset is used for determining a value of the parameter. The second set of data obtained for each of the objects is used as input in a machine learning algorithm and its associated reference value is used as output target.