Extended Brain Sources Estimation via Unrolled Optimization Neural Network

Meng Jiao, Xiaochen Xian, Georges A. Ghacibeh, Feng LiubioRxiv 2022 · DOI

TL;DR

This paper introduces a novel deep learning approach, Unrolled Optimization Neural Network (UONN), to tackle the challenging problem of estimating brain activity sources from EEG/MEG data. By learning a data-driven regularization instead of relying on hand-crafted assumptions, UONN significantly improves the accuracy and efficiency of source localization, particularly for extended brain activity patterns and in noisy conditions.

What Problem Does This Paper Address?

Understanding brain function relies on pinpointing the electrical activity originating from neurons. Electroencephalography (EEG) and Magnetoencephalography (MEG) are powerful tools for measuring this activity with high temporal resolution. However, the signals recorded on the scalp are a "blurry" representation of the underlying brain activity, making it difficult to determine the precise location and extent of the active sources. This is known as the "inverse problem" in EEG/MEG source imaging (ESI).

The core challenge is that the number of potential brain sources far exceeds the number of sensors, leading to an ill-posed problem with infinitely many possible solutions. To overcome this, researchers traditionally employ regularization techniques. These methods introduce assumptions about the nature of brain activity (e.g., that it's sparse or localized) to guide the solution towards a unique and meaningful estimate. However, these hand-crafted regularizations might not perfectly capture the complex spatio-temporal dynamics of brain activity, and the iterative algorithms used to solve them can be computationally expensive.

Methods

The paper proposes a novel framework called Unrolled Optimization Neural Network (UONN) to address the limitations of traditional ESI methods. UONN is built upon the idea of "unrolling" an iterative optimization algorithm, specifically the Iterative Shrinkage-Thresholding Algorithm (ISTA), into a deep neural network.

Here's a breakdown of the key methodological aspects:

  • Unrolling Optimization: Instead of running a traditional iterative algorithm multiple times, UONN unfolds these iterations into distinct layers or "phases" of a neural network. Each phase of the network corresponds to one step of the optimization process.
  • Data-Driven Regularization: A crucial innovation is replacing hand-crafted regularization terms (like Total Variation, which assumes localized spatial gradients) with learnable neural network modules. Specifically, the paper introduces a learnable implicit inverse function, denoted as $\Phi(S)$, which learns the spatial structure of source extents directly from the data. This allows for more flexible and accurate modeling of brain activity patterns.
  • Hyperparameter Learning: Traditional iterative methods often require manual tuning of hyperparameters (like step sizes). UONN learns these hyperparameters automatically during the training process, further optimizing the model's performance.
  • Smoothness Constraint: A smoothness loss ($L_s$) is incorporated into the overall loss function. This loss encourages consistency between successive iterations (phases) of the UONN, mimicking the gradual convergence of iterative algorithms and improving the robustness of the reconstructed source distribution.

The UONN architecture consists of two main modules per phase: an R-module (handling the gradient descent update) and an S-module (handling the proximal mapping with the learned regularization).

Datasets

The paper utilizes both synthetic and real-world datasets for training and evaluation:

  • Synthetic EEG Data: The authors generated synthetic EEG data by simulating brain activity from known source locations. This allowed them to control parameters like the Signal-to-Noise Ratio (SNR) and the spatial extent of the activated sources (using 1, 2, and 3 levels of neighbors). This synthetic data was split into 80% for training and 20% for testing.
  • Real Data (Epilepsy Patients): A publicly accessible dataset from a patient with focal epilepsy was used for validation. This dataset was recorded during a follow-up period after surgery. The researchers used averaged interictal spikes from EEG measurements to localize the epileptogenic zone.

Key Insights and Ideas

  • Learning Spatial Structure is Key: The most significant insight is that learning the spatial structure of brain activity directly from data, rather than imposing predefined assumptions, leads to superior source localization. The learned implicit inverse function $\Phi(S)$ effectively captures complex source extents.
  • Unrolling for Efficiency and Accuracy: By unrolling the optimization process into a deep network, the authors achieve a balance between computational efficiency (replacing slow iterations with fast forward passes) and accuracy, while also enabling data-driven learning.
  • Robustness to Noise and Extended Sources: The UONN method demonstrates remarkable robustness to noise (low SNR) and excels at localizing extended brain activity patterns, areas where traditional methods often struggle.
  • Smoothness as a Regularizer: The inclusion of a smoothness loss term is a clever way to guide the learning process, ensuring that the reconstructed sources evolve in a stable and meaningful manner, similar to how iterative algorithms converge.

Key Results

The UONN method demonstrated significant improvements over benchmark algorithms like MNE, dSPM, and sLORETA:

  • Improved Localization Error (LE) and Area Under Curve (AUC): In synthetic data experiments, UONN consistently achieved lower LE (closer to 0) and higher AUC (closer to 1) across various SNR levels and source extents. For instance, even at low SNRs (e.g., 10 dB and 0 dB), UONN maintained high AUC values (above 0.94) and low LE, outperforming traditional methods whose performance degraded significantly.
  • Superior Performance with Extended Sources: The paper highlights that UONN's advantage is particularly pronounced when dealing with larger, more extended brain activity patterns, where traditional methods tend to produce broader and less precise estimates.
  • Accurate Epilepsy Localization: On real epilepsy data, UONN accurately identified the epileptogenic zone, which was validated by clinical outcomes. The reconstructed sources were cleaner and more precise compared to MNE, clearly delineating the lesion area and differentiating signal strengths in adjacent regions.

Limitations and Future Work

  • Computational Cost of Training: While the UONN offers efficient inference, the training process, involving unrolling an optimization algorithm, can still be computationally intensive.
  • Generalizability to Other Modalities: The current work focuses on EEG/MEG. Future work could explore extending this UONN framework to other neuroimaging modalities or even other inverse problems in science and engineering.
  • Further Exploration of Network Architecture: The paper mentions the network architecture for $\Phi(\cdot)$ and $\theta(\cdot)$ is inspired by autoencoders. Further research could investigate alternative neural network architectures for these components to potentially enhance performance.
  • Investigating Different Optimization Algorithms: The current UONN is based on ISTA. Exploring the unrolling of other optimization algorithms could lead to new insights and improved performance for different types of inverse problems.

Summary updated

Abstract

Abstract Electroencephalography (EEG)/Magnetoencephalography (MEG) source imaging aims to seek an estimation of underlying activated brain sources to explain the observed EEG/MEG recording. Due to the ill-posed nature of inverse problem, solving EEG/MEG Source Imaging (ESI) requires design of regularization or prior terms to guarantee a unique solution. Traditionally, the design of regularization terms is based on preliminary assumptions on the spatio-temporal structure in the source space. In this paper, we propose a novel paradigm to solve the ESI problem by using Unrolled Optimization Neural Network (UONN) (1) to improve the efficiency compared to traditional iterative algorithms; (2) to establish a data-driven way to model the source solution structure instead of using hand-crafted regularizations; (3) to learn the hyperparameter automatically in a data-driven manner. The proposed framework is based on unfolding of the iterative optimization algorithm with neural network modules. The proposed new learning framework is the first one that use the unrolled optimization neural network to solve the ESI problem. The newly designed framework can effectively learn the source extents pattern and achieved significantly improved performance compared to benchmark algorithms.
BibTeX
@misc{jiao2022extended,
  author = {Meng Jiao and Xiaochen Xian and Georges A. Ghacibeh and Feng Liu},
  title = {{Extended Brain Sources Estimation via Unrolled Optimization Neural Network}},
  year = {2022},
  doi = {10.1101/2022.04.11.487935},
  howpublished = {bioRxiv},
}
Figure