[Paper Review] Electromagnetic Source Imaging via a Data-Synthesis-Based Denoising Autoencoder.

Gexin Huang, Zhu Liang Yu+6 more|arXiv (Cornell University)|Oct 24, 2020

Ultrasonics and Acoustic Wave Propagation55 references4 citations

TL;DR

This paper proposes a data-synthesis-based denoising autoencoder (DST-DAE) for electromagnetic source imaging (ESI), leveraging neural networks to learn inverse mappings from measured MEG/EHG signals to cortical sources. By synthesizing large-scale spatio-temporal data using forward models and training a DAE with spatial-temporal feature blocks, the method achieves superior source estimation accuracy and noise robustness compared to traditional prior-based approaches.

ABSTRACT

Electromagnetic source imaging (ESI) is a highly ill-posed inverse problem. To find a unique solution, traditional ESI methods impose a variety of priors that may not reflect the actual source properties. Such limitations of traditional ESI methods hinder their further applications. Inspired by deep learning approaches, a novel data-synthesized spatio-temporal denoising autoencoder method (DST-DAE) method was proposed to solve the ESI inverse problem. Unlike the traditional methods, we utilize a neural network to directly seek generalized mapping from the measured E/MEG signals to the cortical sources. A novel data synthesis strategy is employed by introducing the prior information of sources to the generated large-scale samples using the forward model of ESI. All the generated data are used to drive the neural network to automatically learn inverse mapping. To achieve better estimation performance, a denoising autoencoder (DAE) architecture with spatio-temporal feature extraction blocks is designed. Compared with the traditional methods, we show (1) that the novel deep learning approach provides an effective and easy-to-apply way to solve the ESI problem, that (2) compared to traditional methods, DST-DAE with the data synthesis strategy can better consider the characteristics of real sources than the mathematical formulation of prior assumptions, and that (3) the specifically designed architecture of DAE can not only provide a better estimation of source signals but also be robust to noise pollution. Extensive numerical experiments show that the proposed method is superior to the traditional knowledge-driven ESI methods.

Motivation & Objective

• To overcome the limitations of traditional ESI methods that rely on potentially inaccurate mathematical priors.
• To develop a data-driven approach that learns the inverse mapping from measured electromagnetic signals to cortical sources without relying on explicit prior assumptions.
• To improve estimation performance by leveraging large-scale, realistic synthetic data generated via the forward model of ESI.
• To enhance robustness to noise through a specifically designed denoising autoencoder architecture with spatio-temporal feature extraction.
• To provide a more generalizable and practical solution for electromagnetic source imaging in clinical and research applications.

Proposed method

• A novel data synthesis strategy is employed to generate large-scale spatio-temporal samples by incorporating source prior information into the forward model of ESI.
• The training data are synthesized to reflect realistic source distributions, enabling the network to learn from diverse, plausible source configurations.
• A denoising autoencoder (DAE) architecture is designed with dedicated spatio-temporal feature extraction blocks to model both spatial and temporal dependencies in the data.
• The DAE is trained to reconstruct clean source estimates from noisy input signals, thereby improving robustness to measurement noise.
• The network learns an end-to-end mapping from measured MEG/EEG signals directly to estimated cortical source distributions.
• The method avoids explicit prior formulations by learning the inverse mapping implicitly from the synthetic data distribution.

Experimental results

Research questions

• RQ1 Can a deep learning-based approach outperform traditional ESI methods that rely on hand-crafted priors in terms of source estimation accuracy?
• RQ2 To what extent does data synthesis using the forward model improve the realism and representativeness of training data for ESI?
• RQ3 How does the integration of spatio-temporal feature extraction in the DAE architecture enhance source estimation performance?
• RQ4 Does the denoising autoencoder component improve robustness to noise in measured signals compared to standard autoencoders or traditional methods?
• RQ5 Can the proposed method generalize well to unseen source configurations without overfitting to specific prior assumptions?

Key findings

• The proposed DST-DAE method achieves superior source estimation accuracy compared to traditional knowledge-driven ESI methods in extensive numerical experiments.
• The data-synthesis strategy enables the model to better reflect actual source characteristics than mathematical formulations of prior assumptions.
• The denoising autoencoder architecture significantly improves robustness to noise pollution in the measured signals.
• The spatio-temporal feature extraction blocks effectively capture complex temporal dynamics and spatial patterns in the source signals.
• The method provides a more generalizable solution by learning from a diverse distribution of synthetic data rather than relying on fixed prior constraints.
• The results demonstrate that deep learning with data synthesis can effectively address the ill-posed nature of the ESI inverse problem.