Abstract
The separation of water and fat signals in MRI is a fundamental challenge, traditionally solved via nonlinear optimization algorithms like T2*-IDEAL. However, these methods are highly sensitive to initialization, often leading to "water/fat swaps." While supervised deep learning offers a solution, it requires extensive labeled datasets. This study evaluates three deep neural network (DNN) strategies: Supervised Training (STD), Unsupervised Training (UTD) using a physical forward model as a loss function, and No-Training (NTD) which optimizes a network for a single dataset. Results demonstrate that all three DNN methods achieve high agreement with gold-standard references while offering greater robustness against local minima.
Introduction
In magnetic resonance imaging (MRI), distinguishing between water and fat protons is essential for accurate diagnosis and tissue characterization. Because water and fat resonate at different frequencies—a phenomenon known as chemical shift—their signals interfere. The T2*-IDEAL (Iterative Decomposition of water and fat with Echo Asymmetry and Least-squares estimation) algorithm is the current standard for separating these signals.
Despite its success, T2*-IDEAL faces a significant hurdle: it is a non-convex optimization problem. Without a precise initial guess for the field map, the algorithm can converge to local minima, resulting in "swaps" where water and fat signals are incorrectly identified. While supervised deep learning has recently been applied to this problem, the requirement for large, high-quality labeled datasets limits its clinical flexibility. This paper explores whether physics-informed unsupervised and "no-training" approaches can overcome these limitations.
Methods
The researchers implemented a U-Net architecture to compare three distinct training paradigms:
- Supervised Training (STD): The network is trained using a large dataset where the "ground truth" labels are provided by successful T2*-IDEAL reconstructions.
- Unsupervised Training (UTD): Instead of labels, the network uses the physical forward model of MRI signal decay as its loss function. By predicting water, fat, and field maps, the network attempts to reconstruct the original signal. The loss is calculated as the difference between the reconstructed signal and the actual input.
- No-Training (NTD): Inspired by the "Deep Image Prior" framework, this method does not involve a training phase on a large cohort. Instead, a randomly initialized network is optimized using backpropagation on a single patient's data. The network's weights are adjusted until the output satisfies the physical forward model for that specific scan.
Experimental Design
The study utilized 43 3D-GRE (Gradient Recalled Echo) scans obtained from 12 healthy volunteers and 19 patients. The patient group included individuals with thalassemia major (iron overload) and polycystic kidney disease, providing a challenging range of signal environments. Data were acquired using two 1.5T GE scanners. For the STD and UTD models, the data were partitioned into training, validation, and testing sets.
Results
All three DNN methods demonstrated high fidelity to the reference T2-IDEAL results when the latter was correctly initialized. Quantitative analysis showed correlation coefficients ($R^2$) $\ge 0.98$ for fat fraction, field maps, and R2 decay maps across all methods.
Crucially, the DNN methods—particularly the NTD approach—showed superior robustness. In scenarios where T2*-IDEAL failed due to poor initialization (e.g., initializing the field map to zero), the NTD method successfully converged to the correct water/fat separation. This suggests that the stochastic gradient descent (SGD) used in neural network optimization is more effective at navigating the non-convex landscape of the water/fat problem than traditional iterative solvers.
Furthermore, the study found that the format of the input data significantly influenced performance. Representing the complex MRI signal as magnitude and phase components allowed the networks to learn spatial features more effectively than using real and imaginary components.
Discussion
The success of the Unsupervised (UTD) and No-Training (NTD) methods highlights a shift toward "physics-informed" machine learning. By embedding the mathematical laws of MRI directly into the network's architecture or loss function, researchers can bypass the need for expensive labeled data.
However, practical challenges remain. While the trained models (STD and UTD) provide near-instantaneous results during inference, the NTD method is computationally intensive, requiring approximately 10,000 iterations per scan. This currently limits its utility in time-sensitive clinical settings. Additionally, the study's reliance on 1.5T GE hardware suggests a need for further validation across different field strengths (e.g., 3T) and manufacturers.
Conclusion
Deep neural networks provide a robust framework for solving the water/fat separation problem in MRI. By leveraging physical forward models, unsupervised and no-training methods can achieve gold-standard accuracy without the risks of local minima or the burden of labeled data collection. This work paves the way for more flexible, physics-integrated AI tools in medical imaging.
Summary updated