Cardiac MRF Reconstruction Using Self-Supervised Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning magnetic resonance fingerprinting techniques require large and costly training datasets, are sensitive to respiratory and cardiac motion, and suffer from time-consuming dictionary generation and pattern matching, especially in cardiac applications, leading to inaccurate parameter estimates and generalizability issues.
Innovation Solution
A self-supervised neural network approach that uses a randomly-initialized artificial neural network trained with a single dataset and a physical model, reconstructing tissue parameter maps without a separate training data set, utilizing a self-consistency loss function to minimize reconstruction errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning techniques are used for magnetic resonance fingerprinting, then parameter maps can be reconstructed, but large and costly training datasets are required and generalizability is poor
Solution Approach 1:
The system uses self-supervised learning where the neural network is trained using only the acquired under-sampled k-space data itself, without requiring separate training datasets. The network learns to reconstruct parameter maps by minimizing a self-consistency loss function that compares reconstructed data with the original acquired data, making the system self-sufficient and eliminating the need for large external training data.
Solution Approach 2:
The invention changes the training paradigm from supervised learning with fixed training datasets to self-supervised learning with dynamically generated training examples. By using the acquired data itself as the training source and employing a self-consistency loss function, the system adapts to different patients and scanning conditions without requiring retraining, thus improving generalizability while reducing data requirements.
2Measurement precision
If dictionary-based magnetic resonance fingerprinting is used, then tissue parameters can be quantified, but dictionary generation and pattern matching are time-consuming and memory-intensive
Solution Approach 1:
The invention replaces the traditional dictionary-based pattern matching mechanism with a neural network-based direct reconstruction approach. Instead of generating dictionaries and performing computationally intensive pattern matching, the neural network directly reconstructs parameter maps from under-sampled k-space data, significantly reducing computation time and memory usage while maintaining quantification accuracy.
Solution Approach 2:
The invention extracts and eliminates the time-consuming dictionary generation and pattern matching steps from the MRF pipeline. By using a neural network to directly map k-space data to parameter maps, the system removes the intermediate dictionary-based processing stage, thereby reducing computational burden while preserving the essential tissue parameter quantification function.
3Productivity
If supervised machine learning is used to train neural networks on human subject data, then reconstruction can be performed, but long scan times are required and artifacts occur when features are under-represented
Solution Approach 1:
The system employs self-supervised learning where the neural network trains itself using the acquired under-sampled data without requiring separate fully-sampled reference images. This eliminates the need for long scan times to collect training data, as the network learns from the actual patient data being reconstructed, thereby reducing scan time while maintaining reconstruction capability.
Solution Approach 2:
The invention implements a self-consistency loss function that provides feedback during training by comparing the reconstructed k-space data with the originally acquired k-space data. This feedback mechanism allows the network to learn from the actual data distribution without requiring external training datasets, reducing scan time while preventing artifacts that occur when training data does not represent the actual patient data.
4Measurement precision
If conventional deep learning techniques are used, then reconstruction can be performed, but the system lacks generalizability when patient features or acquisition parameters change
Solution Approach 1:
The invention enables the system to adapt to different patients and acquisition protocols by using self-supervised learning on the actual acquired data. The neural network is trained specifically for each patient's data characteristics without requiring retraining on separate datasets, allowing the system to maintain high reconstruction accuracy across different patients, pathologies, and scanning conditions while improving generalizability.
Data Source
AI summary
A computing system for self-training of a magnetic resonance imaging (MRI) tissue property artificial neural network (ANN) includes a processor and an ANN training application including instructions that, when executed by the one or more processors, are configured to cause the processor to generate tissue property maps; generate MRI fingerprint images; generate reconstructed MRI k-space data; and compare the reconstructed MRI k-space data to acquired MRI k-space data. A computer-implemented method includes generating tissue property maps; generating MRI fingerprint images; generating reconstructed MRI k-space data; and comparing the reconstructed MRI k-space data to acquired MRI k-space data. A non-transitory computer-readable storage medium storing executable instructions that, when executed by a processor, cause a computer to generate tissue property maps; generate MRI fingerprint images; generate reconstructed MRI k-space data; and compare the reconstructed MRI k-space data to acquired MRI k-space data.