Authors: Zhiyong Zhou, Sheng Liao, Xiaolin Li, Youshen Xiao, Bin Dai, Surui Liu, Yakang Dai
Organizations: The School of Biomedical Engineering (Suzhou), Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei 230026, China. · The Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, China. · College of Future Technology, Peking University, Beijing, China.
EEG extended-source imaging aims to localize active cortical regions and characterize their extent, yet supervised inverse networks typically depend on synthetic EEG-source pairs and may be sensitive to synthetic-to-real mismatch. We propose a physics-informed self-supervised implicit neural representation (INR) framework for spatial source reconstruction without paired source labels or simulation-based supervised pretraining. The lead-field matrix encodes the source-to-sensor volume-conduction model and constrains the INR through agreement between predicted and measured EEG. Cortical current density is represented as a continuous coordinate-based neural field. Forward EEG prediction samples the field at cortical vertices and applies local smoothing before lead-field mapping. A classical inverse estimate provides warm initialization and observation-specific spatial guidance through prior-guided support regularization. Simulations across varying SNRs, source extents, and source counts showed low spatial dispersion and good localization performance. Quantitative evaluation on clinical EEG recordings from epilepsy patients and qualitative analysis of scalp EEG during intracranial stimulation supported feasibility for real-world source imaging. These findings suggest the potential of physics-informed self-supervised INR for extended-source imaging.
Figures & tables
Fig. 1: Core concepts of PINR-ESI. (a) Unlike supervised models trained on synthetic EEG–source pairs, PINR-ESI reconstructs sources from measured EEG through forward-model consistency. (b) A continuous coordinate-based neural field represents cortical current density, enabling flexible spatial queries with shared network parameters.
Fig. 2: Overall workflow of PINR-ESI. Source-Map-Guided MLP Warm Start fits a normalized classical inverse estimate to initialize the coordinate-conditioned network. During Forward-Consistency-Guided Self-Supervised Optimization, Forward EEG Prediction (FEP) samples the continuous source field at cortical vertices, applies the fixed local smoothing operator P , and predicts EEG through the lead-field matrix L . Prior-Guided Support Regularization (PGSR) reuses the initial estimate to encourage core activity retention and softly suppress activity outside a broader support region. EEG fitting and PGSR jointly update the network parameters. The optimized field is read out as smoothed source coefficients for reconstruction. The upper-left panel shows the MLP architecture, and the upper-right panel details FEP.
Fig. 3: ESI results of seven methods at SNRs of 5 dB (top) and 20 dB (bottom), with identical single-source ground truth of 10 cm 2 . Both examples share the same single-source ground truth with a target area of 10 cm 2 . Source maps are independently normalized to 0–1. AUC, DLE (mm), SD (mm), and SE (full-map NRMSE) are shown below each reconstruction.
Fig. 4: Performance metrics for seven ESI methods at SNRs of 0, 5, 10, and 20 dB. Bars show means across 100 Monte Carlo trials per condition, with upper error bars indicating one SEM. Each condition includes mixed single- and two-source configurations. DLE and SD are in millimeters. Asterisks indicate superiority of the proposed method using one-sided paired Wilcoxon tests with Holm correction: * p <0.05, ** p <0.01, and *** p <0.001.
100 Monte Carlo trials per condition; same cases used for all methods
Metrics and tests
AUC and NRMSE (dimensionless); DLE and SD (mm); one-sided paired Wilcoxon signed-rank tests with Holm correction
TABLE I: Main evaluation protocol
Fig. 5: ESI results of seven methods with target source areas of 3 cm 2 (top) and 10 cm 2 (bottom) at 10 dB SNR. Source maps are independently normalized to 0–1. The two single-source examples have different source locations.
Fig. 6: Performance metrics for seven ESI methods across target source areas of 3, 5, and 10 cm 2 . Bars show means across 100 Monte Carlo trials per condition, with upper error bars indicating one SEM. DLE and SD are in millimeters. Asterisks indicate superiority of the proposed method using one-sided paired Wilcoxon tests with Holm correction: * p <0.05, ** p <0.01, and *** p <0.001.
Fig. 7: ESI results of seven methods with three (top) and four (bottom) source patches at 10 dB SNR. Each source map is independently normalized to 0–1. Values below each reconstruction indicate AUC, DLE (mm), SD (mm), and SE (energy-aligned full-map NRMSE).
Fig. 8: Performance metrics for seven ESI methods with one to four sources. Bars show means across 100 Monte Carlo trials per condition, with upper error bars indicating one SEM. DLE and SD are in millimeters. Asterisks indicate superiority of the proposed method using one-sided paired Wilcoxon tests with Holm correction: * p <0.05, ** p <0.01, and *** p <0.001.
Fig. 9: Comparison of the complete PINR-ESI reconstruction and the eLORETA estimate used for initialization at SNRs of 0, 5, 10, and 15 dB. Curves show means across 100 source configurations per SNR, with matched cases for both methods. The four panels show AUC, DLE, SD, and NRMSE. DLE and SD are in millimeters.
Fig. 10: Real-data evaluation on the HDEEG-IED-SurgOutcome cohort. (a) EEG and source maps for sub-pat02 at mid-GFP, shown with the RAMPS resection mask. (b) Bidirectional DLE between estimated active cortical regions and the full surface-projected resection region across 24 ILAE-1 subjects. Boxes show medians and interquartile ranges; whiskers follow the 1.5-IQR rule, and diamonds denote means. Outlier symbols are omitted, but all values are included in the statistics.
Fig. 11: Qualitative source imaging results for Localize-MI subjects 1 and 3 in the upper and lower rows, respectively. The insets show the two stimulating contacts as red and yellow points, with their midpoint marked in blue. The projected midpoint is also marked in green on each reconstructed source map. Maps are independently normalized to 0–1 and projected onto the cortical view.
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
Electrophysiological source imaging (ESI) aims to estimate cortical source activity from noninvasive electrophysiological measurements such as electroencephalogram (EEG). However, ESI is fundamentally ill-posed because source activity is substantially higher-dimensional than sensor observations, resulting in non-unique solutions. Recent learning-based approaches address this ambiguity by learning data-driven source priors, yet they often struggle to generalize across subject-specific cortical geometries. To address this, we propose CANDLE, a learning-based ESI model that estimates source activity on subject-specific cortical geometries. CANDLE learns a prior over the null space induced by the source-to-sensor mapping derived from T1-weighted MRI, restricting learning to unobservable source components while preserving geometric constraints. To train CANDLE, we develop a whole-brain simulator spanning over 1,100 subject-specific cortical geometries with source configurations derived from over 26,000 statistical brain maps. Trained exclusively on simulated data, CANDLE outperformed prior ESI methods on simulated source activity estimation and generalized to two empirical tasks: (i) intracranial stimulation localization from simultaneously recorded scalp EEG and (ii) epileptogenic zone estimation from presurgical interictal EEG. Our project page is available at https://candle-esi.pages.dev}{https://candle-esi.pages.dev.
Electroencephalography (EEG) offers millisecond temporal resolution, but inferring underlying neural sources is a severely ill-posed spatial inverse problem. While deep learning has advanced spatial reconstruction, current architectures face a critical dilemma: frame-by-frame models discard vital temporal context, whereas full 4D spatiotemporal networks introduce an architectural trade-off between reconstruction accuracy and inference cost. We propose a novel two-stream framework that explicitly decouples global temporal representation learning from per-time-point spatial refinement. A Transformer-based Temporal Condition Encoder processes the entire EEG sequence via factorized spatiotemporal attention, retaining sensor-resolved features. A fixed inverse then maps these features into source-indexed conditioning for a per-timestep Source-Space Transformer or volumetric convolutional refiner. Extensive evaluations on realistic synthetic data demonstrate that this temporal prior dramatically improves spatial localization, outperforming classical and spatiotemporal baselines, particularly in high-noise and multi-source regimes. Training across diverse leadfields and explicit operator mismatches improves transfer to unseen head geometries and brings template-based reconstruction closer to subject-specific inversion. Furthermore, we apply the model trained only on synthetic EEG data to real-world EEG. A logistic regressor fit on source power differences in eyes-open, eyes-closed conditions successfully decodes age groups.
Marco Morik, Jesse Palarus, Carmen Vidaurre +2
Berlin Institute for the Foundations of Learning and Data, Germany · Machine Learning Group, Technische Universität Berlin, Germany · Basque Center on Cognition, Brain and Language, Mikeletegi Pasealekua, Spain +4
Electroencephalography (EEG) records mixtures of brain-source activity. Even with a known anatomical forward model, experiments that excite only part of the source-state space leave the dynamics unidentified, and repetition cannot resolve the ambiguity. We show that unknown local mechanism changes can supply the missing information. We consider linear dynamics among fixed anatomical sources with known source-state initialization patterns. Changing one source's update rule for one transition leaves a rank-one, source-specific signature in subsequent EEG: subtracting matched baseline responses isolates it, and the forward model identifies the source and calibrates its response history. Combining these histories with initialization responses recovers source interactions without baseline reachability and without first identifying the intervention coefficients. We establish sufficient recovery conditions, a direct estimator, and a noise-sensitivity bound conditional on correct source labels. Simulated EEG on anatomy derived from magnetic resonance imaging confirms the information gain: with baseline excitation confined to four of twelve source coordinates, eight unknown changes recover all dynamics in 32/32 systems, whereas baseline realization, baseline regression through an invertible forward model, and changes that leave the tested states unexposed all fail, and explicitly constructed alternative dynamics reproduce every baseline mean. Where baseline information suffices, direct reconstruction is also more reliable than a matched-information spectral estimator. Nonlocal changes and forward-model error limit accuracy even when source labels are correct.
Ayana Mussabayeva, Jiaqi Sun, Anuar Aimoldin +2
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE · Carnegie Mellon University (CMU), Pittsburgh, PA, USA