DST-CedNet

Electromagnetic Source Imaging via a Data-Synthesis-Based Convolutional Encoder–Decoder Network

Gexin Huang · Ke Liu · Jiawen Liang · Chang Cai · Zheng Hui Gu · Feifei Qi · Yuanqing Li · Zhu Liang Yu · Wei Wu

IEEE Transactions on Neural Networks and Learning Systems, 35(5), 6423–6437, 2024

DST-CedNet learns to reconstruct cortical source time courses from EEG/MEG measurements using synthesized training data. Spatial activation patterns and temporal basis functions generate source signals; a lead field maps them to sensor measurements. A spatiotemporal convolutional encoder–decoder is trained with both source reconstruction and forward-projected sensor reconstruction objectives.

Data synthesis and training overview. Original figure PDF.

Installation

Use Python 3.10 or 3.11 and TensorFlow 2.15. MATLAB Engine is used for data synthesis and the MATLAB comparison methods.

git clone https://github.com/gexinh/DST-CedNet.git
cd DST-CedNet
python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[train]'

For a CPU-only installation, use .[cpu] instead of .[train]. GPU execution additionally requires the CUDA/cuDNN versions supported by your TensorFlow installation.

Install the Engine package matching your MATLAB release. For MATLAB R2023b:

python -m pip install matlabengine==23.2.3

See the MATLAB guide for other releases and temporal basis extraction.

Quick start

Generate training data from the supplied cortical geometry, lead field, and temporal basis functions. The example uses 1024 sources, 63 sensors, and 40 time points. See the data guide to use another head model.

First run a small workflow:

# MATLAB Engine: generate independent train/validation/test splits.
python scripts/synthesize.py --config configs/smoke.yaml

# Python: one epoch and one training batch.
python scripts/train.py --config configs/smoke.yaml

# Predict from the stored test observations.
python scripts/predict.py \
  --weights outputs/smoke_training/best.weights.h5 \
  --manifest outputs/smoke_synthetic/manifest.json \
  --output outputs/smoke_prediction.npz

python scripts/evaluate.py \
  --data outputs/smoke_synthetic/manifest.json \
  --prediction outputs/smoke_prediction.npz \
  --output outputs/smoke_evaluation

The quick start uses four examples per split and one training batch. To train with 27,000 training examples, 27,000 validation examples, and 250 epochs:

python scripts/synthesize.py --config configs/paper.yaml
python scripts/train.py --config configs/paper.yaml

Use configs/paper_two_sources.yaml for two-source training. See the model guide for the architecture and objective.

Evaluation

Compare source-imaging methods on the same synthesized observations. The evaluation includes sensor SNR, source extent, paired source extents, and temporal correlation experiments.

Method	Command name	Runtime	Reference
Minimum-norm estimation (MNE)	mne	Python	Hämäläinen & Ilmoniemi, 1994
Weighted MNE (wMNE)	wmne, wmne-matlab	Python / MATLAB	Dale & Sereno, 1993
LORETA	loreta	MATLAB	Pascual-Marqui et al., 1994
Sparse Bayesian learning (SBL)	sbl	MATLAB	Wipf & Rao, 2004, Wipf & Nagarajan, 2009
FOCUSS	focuss	MATLAB	Gorodnitsky et al., 1995
STTONNICA	sttonnica	MATLAB + STTONNICA package	Valdés-Sosa et al., 2009

For a small comparison:

python scripts/run_experiments.py --run --generate \
  --synthesis-config configs/synthesis.yaml \
  --condition snr_-5 --samples 2 \
  --methods wmne-matlab loreta sbl focuss \
  --output outputs/comparison

Method parameters are in configs/baselines.yaml. Results contain source estimates, per-trial metrics, and mean/SEM summaries for each method. STTONNICA uses its author's MATLAB package via --sttonnica-dir; see setup and references. The evaluation guide covers DST-CedNet comparisons, all 16 conditions, training-size experiments, and metric definitions.

Network architecture figure

Original figure PDF.

Citation

@article{huang2024dstcednet,
  author  = {Huang, Gexin and Liu, Ke and Liang, Jiawen and Cai, Chang and Gu, Zheng Hui and Qi, Feifei and Li, Yuanqing and Yu, Zhu Liang and Wu, Wei},
  title   = {Electromagnetic Source Imaging via a Data-Synthesis-Based Convolutional Encoder--Decoder Network},
  journal = {IEEE Transactions on Neural Networks and Learning Systems},
  volume  = {35},
  number  = {5},
  pages   = {6423--6437},
  year    = {2024},
  doi     = {10.1109/TNNLS.2022.3209925}
}

License and sources

Original implementation code is provided under the MIT License. Paper figures, modeling assets, and external software have separate sources and terms; see source attribution.