Better than real: Complex-valued neural nets for MRI fingerprinting

Patrick Virtue, Stella X. Yu, Michael Lustig

TL;DR

This paper introduces a deep learning approach for MRI fingerprinting, moving beyond slow dictionary-based methods. They developed a novel complex-valued neural network with a unique "cardioid" activation function, demonstrating significantly improved accuracy and efficiency in mapping MRI signals to tissue parameters compared to existing real-valued neural networks.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) fingerprinting aims to precisely determine tissue properties, like T1 and T2 relaxation times, from complex-valued MRI signals. These parameters are crucial for distinguishing between different tissue types (e.g., gray vs. white matter) and identifying pathologies. The traditional method, "dictionary-based" fingerprinting, involves comparing incoming MRI signals to a pre-computed library of signals associated with known tissue parameters. However, this approach becomes computationally prohibitive as the number of tissue parameters or the desired precision increases, leading to slow performance and scalability issues. The paper addresses this by proposing a deep learning solution that learns a direct, efficient mapping from MRI signals to tissue parameters.

Methods

The core of this paper's contribution lies in two novel aspects:

  1. Deep Learning for Inverse Mapping: Instead of relying on pre-computed dictionaries, the authors train a deep neural network to directly learn the inverse mapping from MRI signals to tissue parameters. This is achieved by generating synthetic MRI data using a Bloch simulator, paired with ground-truth tissue parameters (T1, T2, and B0). This synthetic dataset allows for efficient training of the neural network.
  2. Complex-Valued Neural Networks with Cardioid Activation: Recognizing that MRI signals are inherently complex-valued, the authors propose a complex-valued neural network architecture. A key innovation here is the development of a new complex activation function called the "complex cardioid." This function is phase-sensitive, attenuating the output magnitude based on the input phase, while preserving the phase itself. Crucially, when restricted to real inputs, it behaves like the standard ReLU activation function. This complex-valued approach, combined with the cardioid activation, is hypothesized to better leverage the phase information present in MRI signals, which is lost when treating complex signals as two separate real channels. They also utilize Wirtinger calculus (CR calculus) for gradient descent optimization in their complex networks.

Datasets

The paper uses synthetic MRI data generated from a Bloch simulator. This simulator takes ground-truth tissue parameters (T1, T2, B0) and scanner parameters to produce complex-valued MRI temporal signals. This synthetic dataset is used for training the neural networks.

For evaluation, they use a numerical MRI phantom with known T1, T2, and proton density values, as described in previous work. They also introduce a linear ramp in the B0 field across the image and test with both clean phantom signals and signals corrupted by complex-valued Gaussian noise, achieving a peak signal-to-noise ratio (pSNR) of 40.

Key Insights and Ideas

  • Complex Signals Demand Complex Networks: Treating complex MRI signals as two independent real channels (real and imaginary parts) is suboptimal. Complex-valued neural networks, by operating directly on complex numbers and respecting phase information, can achieve superior performance.
  • The Cardioid Activation is Key: The novel complex cardioid activation function is a significant enabler. It allows complex networks to learn effectively by intelligently attenuating signal magnitudes based on phase, outperforming other complex activation functions like separable sigmoid and siglog.
  • Deep Learning Scales Better: Deep learning offers a computationally efficient alternative to dictionary-based methods, with training time that doesn't scale with the complexity of the parameter space, making it more suitable for higher-dimensional problems.

Key Results

  • Accuracy Improvement: Complex-valued neural networks, particularly those using the cardioid activation, significantly outperform both the traditional nearest neighbor baseline and real-valued neural networks in terms of Normalized Root Mean Square Error (NRMSE) for T1 and T2 parameter estimation, especially under noisy conditions. For example, with clean signals, the complex cardioid network achieved an NRMSE of 1.42 for T1, compared to 10.63 for nearest neighbor.
  • Efficiency: The paper highlights the computational advantage of neural networks. While dictionary-based methods' computational cost explodes exponentially with more parameters, neural networks offer a more manageable increase in computational cost (FLOPs) for predicting more parameters.
  • B0 Prediction: While complex networks excel at T1 and T2, the traditional nearest neighbor approach showed slightly better accuracy for B0 prediction in their experiments.

Limitations and Future Work

  • B1 Inhomogeneity: The current experiments do not include B1 inhomogeneity in the simulations. The authors suggest that a fourth neural network could be added to incorporate this or other parameters.
  • Pixel-wise vs. Full Image: The current work focuses on pixel-wise fingerprinting reconstruction. The authors plan to extend their approach to full image predictions for under-sampled MRI fingerprinting in the future.
  • Generalizability: While the synthetic data generation is a strength, further validation on real-world scanner data would be beneficial.

Summary updated

Abstract

The task of MRI fingerprinting is to identify tissue parameters from complex-valued MRI signals. The prevalent approach is dictionary based, where a test MRI signal is compared to stored MRI signals with known tissue parameters and the most similar signals and tissue parameters retrieved. Such an approach does not scale with the number of parameters and is rather slow when the tissue parameter space is large. Our first novel contribution is to use deep learning as an efficient nonlinear inverse mapping approach. We generate synthetic (tissue, MRI) data from an MRI simulator, and use them to train a deep net to map the MRI signal to the tissue parameters directly. Our second novel contribution is to develop a complex-valued neural network with new cardioid activation functions. Our results demonstrate that complex-valued neural nets could be much more accurate than real-valued neural nets at complex-valued MRI fingerprinting.
BibTeX
@inproceedings{virtue2017better,
  author = {Patrick Virtue and Stella X. Yu and Michael Lustig},
  title = {{Better than real: Complex-valued neural nets for MRI fingerprinting}},
  booktitle = {ICIP},
  year = {2017},
  url = {https://openreview.net/forum?id=-1FelMayBr},
}
Figure