acceptodds
Under review as a conference paper at ICLR 2027

CANDLE: Cortical Null-Space Decomposition for Noninvasive Brain Source Imaging

Abstract

Electrophysiological source imaging (ESI) aims to estimate cortical source activity from noninvasive electrophysiological measurements such as EEG. However, ESI is fundamentally ill-posed because source activity is substantially higher-dimensional than sensor observations, resulting in non-unique solutions. Conventional methods resolve this ambiguity using hand-designed constraints, while recent learning-based approaches instead learn data-driven source priors. However, the latter often struggle to generalize across individual cortical geometries. To this end, 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 T1-derived source-to-sensor mapping, restricting learning to unobservable source components while preserving geometric constraints. To train CANDLE, we developed 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 estimation and generalized to two empirical tasks: (i) intracranial stimulation-site localization and (ii) epileptogenic zone estimation from presurgical interictal EEG.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography (EEG), Electrophysiological source imaging (ESI), Brain Computer Interface (BCI)

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PYoxOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The CANDLE study estimates individual cortical source activity from EEG. The model constructs a forward matrix L using the individual T1 MRI, then denoises it with the network f_θ, and recovers the observable components with L^† f_θ (X), and supplements the zero-space components with (I - L^† L) g_γ (L^† f_θ (X)). The training loss includes source reconstruction, sensor denoising, and zero-space error (page 3, §2.1, Equations (1)–(3)). The training simulator combines 1,113 individual geometries, 994 cortical regions of the Jansen–Rit neural population model, and 800 source configurations obtained from 26,273 statistical brain maps. The authors report the simulation tests divided by individuals, the localization of 291 stimulation records from 32 patients, and the reference positioning of the resection areas of 22 patients with good postoperative outcomes.

Critical strengths

The combination of physical observation operators and learning priors has clear technical reasons. The individual L directly determines the projection and the analytical inverse, and the network does not need to redefine the output dimension for each new subject; denoising before inversion also addresses the problem of pseudo-inverse amplifying noise. The entire method does not mistakenly attribute the zero-space unobservability to the absence of a priori knowledge for the unique recovery. This provides a reasonable inductive bias for cross-individual applications.

The tests cover both controlled true values and two actual positioning scenarios simultaneously, and the simulation results have meaningful amplitudes. The simulation adopts a subject-wise division, with the Dice of CANDLE being 0.48, while the best control is 0.33, and the HD95 is also the lowest. The real stimulation results are first aggregated within the subjects, reducing the risk of treating the same patient as multiple independent patients; the authors also publicly explain that the overall difference in the stimulation task relative to sLORETA/GBF is not significant, and the significance of the EZ task only covers part of the controls. These results support further research on this scheme and cannot be simply classified as only having simulation verification.

Critical weaknesses

Currently, the contribution of "individualized zero-space design" compared to general learning models has not been fully recognized. When g_γ is removed in Figure 6, the entire learning source prior is also removed, making it impossible to distinguish the effects of zero-space projection and non-restricted source correction with the same capacity. On the other hand, fixing the fsaverage simulator reduces the stimulus peak distance from 32.9 to 33.6 mm; randomly selecting ROIs increases it to 33.4 mm, while improving EZ HD95 from 61.7 to 59.0 mm. The related simulator ablation report does not show a significant advantage (pages 8-9, Figure 6, §3.4). Moreover, the main text and appendix do not specify whether ConvDip/DeepSIF uses the same anatomical simulation training set, budget, and individual source mapping; only stating that all three use simulation data is not sufficient to confirm this (pages 5, 15). Therefore, the existing results prove the effectiveness of the entire system, but it is still difficult to determine whether the gain comes from the geometric/zero-space constraints emphasized in this paper, or the training distribution or backbone differences. This is a lack of key distinguishing evidence, not because the cited general zero-space methods cannot be used for new problems.

The key processing of real data entering the fixed model is not explained, affecting the verifiability of the geometric constraints and zero-sample results. The model is fixed at BioSemi 64 channels, 500 Hz, 1 second length; pre-processing only includes downsampling, filtering, average reference, and z-score (page 15, B.2). However, the original EZ data provides 257/204 channels, different lengths of average discharge, and the cleaned version is usually 250 Hz, [2] while the manuscript does not explain the channel selection or interpolation, short sequence processing, and whether these transformations are applied synchronously to L. The z-score axis and scale used, as well as how the target is denoised and the L target, are also not explained. If the input/noise reduction target is scaled channel by channel while L maintains another set of units, the analytical part of equation (2) cannot be directly interpreted as an inversion in the same physical coordinates; the current materials cannot determine if the authors have correctly handled it. Additionally, the stimulation evaluation focuses on the ±5 ms window, but does not explain whether the original current pseudotrajectory is retained, deleted, or interpolated in this window. These are unclarified information, not an already confirmed implementation error, but directly affects how the main real data evidence should be interpreted.

Please provide the complete mapping of two real datasets to 64 channels, 500 sampling points, and the reference, scale, region direction, and L processing; also explain the de-noising status of the ±5 ms window, the inclusion rules for 22 EZ patients. If there is a script, please explain the results as to whether they do indeed correspond to the observed model of equation (2) in the processing sequence and necessary formulas. This is mainly to clarify the existing experiments. In the saved predictions, supplementing the peak to the closest distance to the resection cavity is also helpful in judging whether the existing HD95/mass center conclusions depend on the cavity size; no new patient data collection is required.

Questions for authors

see Weaknesses

4 comments

We sincerely thank the reviewer for the thorough reading of our manuscript. In response, we have added an ablation isolating the null-space constraint and clarified the baseline training configuration and the full preprocessing pipeline for the empirical datasets. All revisions are highlighted in blue in the revised manuscript; below, W denotes Weaknesses.

We also noticed that the reviewer set "Flag For Ethics Review" to "Yes". To our understanding, neither the review nor the "Details Of Ethics Concerns" field mentions an ethical issue. Our study uses only simulated data and secondary analyses of publicly available datasets, in accordance with their data-use terms and the ethical approvals reported in the original studies (see the Ethics Statement). We would therefore greatly appreciate it if the reviewer could share the specific concern behind the flag, so that we can address it and further improve the transparency of our work.


W1-(i): Disentangling the contribution of the null-space constraint from the learned source prior

To disentangle the advantage of restricting learning to the null space from that of conditioning on subject-specific cortical geometry in CANDLE, we conducted an additional ablation study with Model (b-iii) (see the table below; SPES: intracranial stimulation localization task; EZ: epileptogenic zone estimation task). Model (b-iii) removes the null-space projection LnL_n applied to gγg_\gamma in Equation (2), yielding Y^=L†Xdenoise+gγ(L†Xd)\hat{Y} = L^\dagger X_\mathrm{denoise} + g_\gamma(L^\dagger X_d). Model (b-iii) shares the same architecture, number of parameters, and training configuration as CANDLE (Model (a)). Consequently, any difference between Models (a) and (b-iii) is attributable solely to whether learning is restricted to the null space.

As shown in the table, Model (b-iii) underperformed Model (a) on all four metrics on real data. The degradation was particularly pronounced for EZ estimation, where HD95 and centroid distance increased by 3.75 mm and 2.86 mm, respectively. Without the null-space projection in Model (b-iii), the output of gγg_\gamma may leak into the observable range-space component, potentially producing estimates that are inconsistent with the measurements (i.e., LY^≠XdenoiseL\hat{Y} \neq X_{\text{denoise}}). The null-space constraint prevents this by confining the learned prior to components that are not determined by the observations, which contributes to the improved estimation accuracy. These additional results have been incorporated into Section 3.4 of the revised manuscript.

ModelSPES (Spatial Disp. [mm]) ↓SPES (Peak Dist. [mm]) ↓EZ (HD95 [mm]) ↓EZ (Centroid Dist. [mm]) ↓
(a) CANDLE54.11±1.2532.86±1.9161.72±2.7535.38±2.98
(b-iii) w/o LnL_n proj. on gγg_\gamma54.58±1.2633.76±1.8765.47±2.3038.24±3.13

W1-(ii): Marginal gains from simulator design

As reported in Subsection 3.4 and Figure 6, our simulator ablations show that both (d-i) fixing the head model to fsaverage and (d-ii) randomly selecting ROIs degrade performance relative to the full CANDLE model (a), on four and three of the four empirical metrics, respectively. Although these differences are not statistically significant, they exhibit a consistent trend of degradation across two fundamentally different empirical tasks: intracranial stimulation localization and epileptogenic zone estimation. We therefore suppose that both the diversity of cortical geometries and the realistic ROI definitions in our simulator contribute meaningfully, rather than marginally, to the effective training of CANDLE.


W1-(iii): Configuration of the ConvDip and DeepSIF Baselines

ConvDip [1] and DeepSIF [2] were trained on the same simulated data as CANDLE, using identical preprocessing procedures, training/validation/test splits, and model selection criteria, as detailed in Appendix B.2. Regarding source-space mapping, all learning-based ESI methods predict source activity across 994 cortical regions defined by the Hagmann–Lausanne parcellation [3]. As these regions are defined on each subject's FreeSurfer-reconstructed white-matter surface [4], the predictions are mapped to subject-specific source spaces through atlas label correspondence. However, unlike CANDLE, ConvDip and DeepSIF cannot account for subject-specific cortical geometry, and thus inter-individual variations in regional locations and shapes are not reflected in their estimates. We have clarified these details in L296 and Appendix C.1 of the revised manuscript.

W2 & W3: Details for real-data preprocessing

We describe the detailed evaluation settings for empirical data in the order the points were raised below. For brevity, we refer to the intracranial stimulation localization task as SPES and to the epileptogenic zone estimation task as EZ. Clarifications below have been added to Appendix B.2 of the revised manuscript.

Channel selection (EZ/SPES): The datasets used for the EZ and SPES evaluations include subjects recorded with either 257 or 204 channels. For each recording, we kept only the electrodes whose labels match those of the BioSemi 64-channel montage [5], which is CANDLE’s default montage (Appendix B.2), and used these channels for evaluation.

Variable segment lengths and sampling rate (EZ): As described in Appendix B.2, CANDLE takes as input a 1-s segment sampled at 500 Hz. The EZ data, however, vary in length and were sampled at a lower rate of 250 Hz. We therefore first upsampled the data to 500 Hz. Segments longer than 500 samples were center-cropped and shorter ones were zero-padded to 500 samples. Estimates for the zero-padded intervals were left out of the evaluation.

Unit consistency between the EEG and the lead-field matrix L\mathbf{L}: In all experiments, we applied the same preprocessing to L\mathbf{L} as to the EEG (e.g., common average referencing), so L\mathbf{L} and the EEG remain in consistent units. Please see our response to Reviewer sBb5 (Q3) for details.

Treatment of the ±5 ms window (SPES): Following prior ESI work [6], we applied no preprocessing or artifact removal to this window beyond what is described in Appendix B.2.

Source orientation (EZ/SPES): Throughout this study, we assume that source activity is oriented perpendicular to the cortical surface in each region.

Inclusion criteria for the 22 EZ patients: As stated in Subsection 3.3, we included patients with a favorable outcome 12 months after surgery, defined as ILAE [7] class 1 or 2.

Implementation of the EZ evaluation: The EZ evaluation is implemented in CANDLE/src/datasets/vorderwullbecke_2025/evaluate.py in the supplementary material.

Implementation of Equation (2): Equation (2) is implemented in CANDLE/src/model/CANDLE.py in the supplementary material.

Dependence of the EZ evaluation on resection-cavity size: The centroid distance in the current manuscript is the distance between the centroid of the resection cavity and the centroid of the estimated source activity. It therefore does not depend on the size of the resection cavity.


References

[1] Lukas Hecker, Rebekka Rupprecht, Ludger Tebartz Van Elst, and J ̈urgen Kornmeier. ConvDip: A convolutional neural network for better EEG source imaging. Front. Neurosci., 15:569918, 2021.

[2] Rui Sun, Abbas Sohrabpour, Gregory A Worrell, and Bin He. Deep neural networks constrained by neural mass models improve electrophysiological source imaging of spatiotemporal brain dynamics. PNAS, 119(31):e2201128119, 2022.

[3] Patric Hagmann, Leila Cammoun, Xavier Gigandet, Reto Meuli, Christopher J Honey, Van J Wedeen, and Olaf Sporns. Mapping the structural core of human cerebral cortex. PLoS biology, 6(7):e159, 2008.

[4] Bruce Fischl. FreeSurfer. NeuroImage, 62(2):774–781, 2012.

[5] Robert Oostenveld et al. The five percent electrode system for high-resolution EEG and ERP measurements. Clin. Neurophysiol., 112(4):713–719, 2001.

[6] Song Wang, Kexin Lou, Chen Wei, Zhiyuan Sheng, Jiahao Tang, Kaining Peng, Xinke Shen, Shuhao Mei, Liang Chen, Dongfeng Gu, et al. A geometry aware framework enhances noninvasive mapping of whole human brain dynamics. Nat. Biomed. Eng., 2026.

[7] HG Wieser, WT Blume, D Fish, E Goldensohn, A Hufnagel, D King, MR Sperling, and H Luders. Proposal for a new classification of outcome with respect to epileptic seizures following epilepsy surgery. Epilepsia, 42(2):282–286, 2001.

The authors have responded to each of my points and narrowed the claims to what the experiments actually support. I confirm that there are no remaining issues, and I have raised my score. Best of luck.

We sincerely appreciate your continued engagement and reassessment. While you confirmed that there are no remaining issues, your rating is still "Weak Rejection." Does this mean that you still have some concerns about our work? If so, we would greatly appreciate any additional feedback you could provide to help us further improve our work.

Reviewer_xRnrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CANDLE, a framework for estimating cortical source activity from EEG while accounting for individual differences in cortical geometry. Using a subject-specific lead-field matrix derived from T1-weighted MRI, the method combines EEG denoising with range-null space decomposition to reconstruct the observable source component and learn a prior for the unobservable component. The paper also develops a whole-brain simulator incorporating anatomical geometries from over 1,100 individuals and source configurations derived from over 26,000 statistical brain maps. Trained exclusively on simulated data, CANDLE is evaluated on both simulated source reconstruction and real-world tasks: intracranial stimulation localization and epileptogenic zone estimation. The results show improved overall spatial localization performance compared with baselines.

Critical strengths
  • The paper is clearly written and well organized. The motivation and main ideas are easy to understand, and the figures help explain the approach.
  • The evaluation is broad, covering simulated data and two real-world tasks, with comparisons against several established baselines. CANDLE achieves strong overall performance, supported by both quantitative results and qualitative examples. Applying the model to real recordings without subject-specific fine-tuning is also interesting.
  • The design of main components of the overall pipeline looks well motivated.
Critical weaknesses
  • The quantitative evaluation focuses mainly on spatial localization, while temporal reconstruction is supported mostly by qualitative examples. It would be useful to assess how accurately it recovers their temporal dynamics, e.g., source-level temporal correlation, spectral similarity, phase, functional connectivity, etc.
  • Minor: Though the choice of each module design is well motivated, the methodological contribution appears primarily to be the integration of existing null-space learning techniques with subject-specific EEG forward models and a large-scale simulator. The empirical value of this integration is promising, but the current ablations do not fully establish the benefits of null-space constraint, e.g., removing the estimator g also removes the entire learned source-space correction module
Questions for authors
  1. I was wondering if the author have explored extending CANDLE to a voxel-based source space? How sensitive is CANDLE's localization performance to the granularity of the cortical parcellation?
  2. Have the author explored how CANDLE could recover temporal dynamic related features?
  3. Have the author evaluated how CANDLE’s localization performance changes as the number of EEG input channels is reduced, for example from 64 to 32 or 16? It would be useful to understand when performance starts to decrease substantially and whether you can recommend a minimum channel count for using/adapting this proposed approach
4 comments

We sincerely thank the reviewer for the positive assessment and constructive feedback. In response, we have added a quantitative evaluation of temporal dynamics, an ablation isolating the null-space constraint, and an analysis of sensitivity to sensor density. All revisions are highlighted in blue in the revised manuscript; below, W and Q denote Weaknesses and Questions, respectively.


W1 & Q2: Quantitative evaluation of temporal dynamics

As noted in L306, spatial localization is the primary objective of ESI, and we therefore adopted the evaluation metrics established in prior studies (e.g., HD95 and Dice score) [1,2]. Nevertheless, we agree that temporal reconstruction should also be assessed quantitatively, and have therefore evaluated the correlation between the functional connectivity (FC) matrices derived from the estimated and ground-truth source activity (FC correlation) in the simulated source activity estimation. FC summarizes the temporal correlations among cortical regions, and computing it separately for the broadband (1–40 Hz) and the δ (1–4 Hz), θ (4–8 Hz), and α (8–13 Hz) bands further enables the assessment of frequency-specific temporal dynamics. These three bands were chosen because the neural dynamics generated by the Jansen–Rit model underlying our simulator (Section 2.2) is predominantly concentrated within them. As shown in the table below, CANDLE achieved the highest FC correlation across all four bands, with statistically significant improvements over all baselines in the broadband and δ band (p<0.01p<0.01).

Because ground-truth source activity is unavailable for empirical data, FC correlation was evaluated only on simulated data; for an indirect validation of temporal dynamics on empirical data, please refer to our response to Reviewer sBb5 (W2). The FC evaluation has been incorporated into Section 3.1, Figure 3, and Appendix C.2 of the revised manuscript.

MetricLCMVsLORETAwMNEγ-MAPChampagneeLORETAConvDipDeepSIFGBFCANDLE
FC Corr. (1–40 Hz) ↑0.06±0.00††0.06 \pm 0.00^{\dagger\dagger}0.11±0.00††0.11 \pm 0.00^{\dagger\dagger}0.11±0.00††0.11 \pm 0.00^{\dagger\dagger}0.08±0.00††0.08 \pm 0.00^{\dagger\dagger}0.08±0.00††0.08 \pm 0.00^{\dagger\dagger}0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.20‾±0.00††\underline{0.20} \pm 0.00^{\dagger\dagger}0.17±0.00††0.17 \pm 0.00^{\dagger\dagger}0.15±0.00††0.15 \pm 0.00^{\dagger\dagger}0.26±0.00\mathbf{0.26 \pm 0.00}
FC Corr. (δ\delta, 1–4 Hz) ↑0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.08±0.00††0.08 \pm 0.00^{\dagger\dagger}0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.06±0.00††0.06 \pm 0.00^{\dagger\dagger}0.27‾±0.00‡\underline{0.27} \pm 0.00^{\ddagger}0.21±0.00††0.21 \pm 0.00^{\dagger\dagger}0.17±0.00††0.17 \pm 0.00^{\dagger\dagger}0.31±0.00\mathbf{0.31 \pm 0.00}
FC Corr. (θ\theta, 4–8 Hz) ↑0.01±0.00††0.01 \pm 0.00^{\dagger\dagger}0.09±0.000.09 \pm 0.000.10±0.000.10 \pm 0.000.07±0.000.07 \pm 0.000.06±0.000.06 \pm 0.000.10‾±0.00\underline{0.10} \pm 0.000.10±0.00††0.10 \pm 0.00^{\dagger\dagger}0.09±0.00††0.09 \pm 0.00^{\dagger\dagger}0.06±0.00††0.06 \pm 0.00^{\dagger\dagger}0.15±0.00\mathbf{0.15 \pm 0.00}
FC Corr. (α\alpha, 8–13 Hz) ↑0.00±0.000.00 \pm 0.000.00±0.000.00 \pm 0.000.00±0.000.00 \pm 0.000.00±0.000.00 \pm 0.00−0.03±0.00-0.03 \pm 0.00−0.01±0.00-0.01 \pm 0.000.02±0.00†0.02 \pm 0.00^{\dagger}0.01±0.00†0.01 \pm 0.00^{\dagger}0.02‾±0.00\underline{0.02} \pm 0.000.05±0.00\mathbf{0.05 \pm 0.00}

W2: Effect of the null-space constraint

As the reviewer pointed out, Model (b-ii) removes the learned source-space correction altogether. We therefore added Model (b-iii) to Subsection 3.4 of the revised manuscript, which keeps gγg_\gamma but omits the null-space projection Ln\mathbf{L}_n on its output, thereby isolating the contribution of the null-space constraint. Detailed results are provided in our response to Reviewer PYox (W1-(i)).

Q1: Extension to a voxel-based source space

The range–null space decomposition and the other modules of CANDLE are independent of how the source space is defined. CANDLE can therefore be extended to a voxel-based source space simply by replacing its region-wise estimates over cortical parcels with voxel-wise estimates. However, scalp EEG predominantly reflects the synchronized activity of cortical pyramidal neurons [3]. We therefore consider the main advantage of a voxel-based source space, namely its ability to capture subcortical activity, to be of limited benefit for EEG-based source imaging.

Regarding parcellation granularity, CANDLE adopts the 994-region Hagmann/Lausanne parcellation [4], in which each region covers approximately 1.5 cm² of cortical surface. As described in Appendix B.1, cortical activity must be synchronized over approximately 6–10 cm² to produce scalp potentials detectable by EEG [5]. The current regions are thus well below this spatial scale. We therefore consider the current granularity to be near the finest level resolvable from EEG on neurophysiological grounds, and further refinement is unlikely to yield substantial improvements.


Q3: Robustness to reduced EEG channel counts

As noted in the Limitations, CANDLE and the other learning-based methods (ConvDip and DeepSIF) are currently trained on a fixed BioSemi 64-channel montage [6], and extending them to other electrode configurations is left for future work. Nonetheless, we agree that reduced channel counts are important in practice, and the table below reports results for simulated source activity estimation (Simulation) and intracranial stimulation localization (SPES) with 64, 32, and 16 electrodes. For the learning-based methods, missing channels were linearly interpolated from neighboring electrodes to recover the 64-channel input. CANDLE consistently outperformed the baselines with fewer electrodes, with significant improvements in Simulation under all conditions (p<0.001p < 0.001). Its own performance dropped most when reducing from 32 to 16 electrodes in both tasks (HD95: +7.23 mm in Simulation), suggesting that the minimum electrode count for practical use lies within this range. Further results, including other baselines and metrics, are provided in Section 3.5 of the revised manuscript.

ExperimentMetric# ElectrodeSecond-best baselineBest baselineCANDLE
SimulationHD95 [mm] ↓6445.12±0.61††45.12 \pm 0.61^{\dagger\dagger} (DeepSIF)41.05±0.57††41.05 \pm 0.57^{\dagger\dagger} (ConvDip)37.26±0.74\mathbf{37.26 \pm 0.74}
SimulationHD95 [mm] ↓3249.16±0.67††49.16 \pm 0.67^{\dagger\dagger} (DeepSIF)45.02±0.58††45.02 \pm 0.58^{\dagger\dagger} (ConvDip)37.41±0.73\mathbf{37.41 \pm 0.73}
SimulationHD95 [mm] ↓1656.67±0.68††56.67 \pm 0.68^{\dagger\dagger} (DeepSIF)50.38±0.60††50.38 \pm 0.60^{\dagger\dagger} (ConvDip)44.64±0.74\mathbf{44.64 \pm 0.74}
SPESPeak distance [mm] ↓6437.47±2.4237.47 \pm 2.42 (sLORETA)36.47±1.9436.47 \pm 1.94 (GBF)32.86±1.91\mathbf{32.86 \pm 1.91}
SPESPeak distance [mm] ↓3243.42±1.82††43.42 \pm 1.82^{\dagger\dagger} (eLORETA)42.96±2.28††42.96 \pm 2.28^{\dagger\dagger} (ConvDip)34.92±1.72\mathbf{34.92 \pm 1.72}
SPESPeak distance [mm] ↓1648.12±2.42††48.12 \pm 2.42^{\dagger\dagger} (DeepSIF)44.01±2.4644.01 \pm 2.46 (GBF)40.18±2.01\mathbf{40.18 \pm 2.01}

References

[1] Rui Sun, Abbas Sohrabpour, Gregory A Worrell, and Bin He. Deep neural networks constrained by neural mass models improve electrophysiological source imaging of spatiotemporal brain dynamics. PNAS, 119(31):e2201128119, 2022.

[2] Song Wang, Kexin Lou, Chen Wei, Zhiyuan Sheng, Jiahao Tang, Kaining Peng, Xinke Shen, Shuhao Mei, Liang Chen, Dongfeng Gu, et al. A geometry aware framework enhances noninvasive mapping of whole human brain dynamics. Nat. Biomed. Eng., 2026.

[3] Hans Hallez, Bart Vanrumste, Roberta Grech, Joseph Muscat, Wim De Clercq, Anneleen Vergult, Yves D’Asseler, Kenneth P Camilleri, Simon G Fabri, Sabine Van Huffel, et al. Review on solving the forward problem in eeg source analysis. J. NeuroEng. Rehabil., 4(1):46, 2007.

[4] Patric Hagmann, Leila Cammoun, Xavier Gigandet, Reto Meuli, Christopher J Honey, Van J Wedeen, and Olaf Sporns. Mapping the structural core of human cerebral cortex. PLoS biology, 6(7):e159, 2008.

[5] James X Tao, Amit Ray, Susan Hawes-Ebersole, and John S Ebersole. Intracranial EEG substrates of scalp EEG interictal spikes. Epilepsia, 46(5):669–676, 2005.

[6] Robert Oostenveld et al. The five percent electrode system for high-resolution EEG and ERP measurements. Clin. Neurophysiol., 112(4):713–719, 2001.

I thank the authors for the thoughtful rebuttal, additional experiments, and clarifications, which address most of my questions and strengthen the manuscript. Overall, I appreciate the authors' effort and I am happy to maintain my positive recommendation for acceptance

We are glad that our responses and additional experiments have addressed your concerns. We sincerely appreciate your constructive feedback and positive assessment of our work. Thank you again for your time and consideration.

Reviewer_sBb5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CANDLE is a learning-based EEG source-imaging method that combines subject-specific forward models with a learned null-space prior. It denoises EEG, analytically reconstructs the observable source component, and estimates the unobservable component using a neural network. Training uses a biophysical simulator spanning 1,113 subject anatomies and source configurations derived from statistical brain maps. Evaluation includes simulated reconstruction, intracranial stimulation localization in 32 patients, and epilepsy-related localization in 22 surgical patients.

Critical strengths
  1. Principled formulation. The range–null-space decomposition provides a clear framework for combining subject-specific anatomical constraints with learned source priors. 2 Clear methodological reporting. Training, preprocessing, and key hyperparameters are well explained, including optimization settings, checkpoint selection, EEG processing, head-model construction, and neural-mass simulation parameters.
  2. Meaningful sim-to-real evaluation. The two real-human datasets provide complementary localization tests beyond synthetic reconstruction.
  3. Useful component analysis. The paper examines the denoiser, null-space estimator, auxiliary losses, anatomical diversity, and source-pattern construction rather than reporting only the full model’s performance.
Critical weaknesses
  1. Incomplete comparison with sparse Bayesian ESI. The literature discussion and experiments omit Champagne and related approaches such as Gamma-MAP. Since improved localization and reduced diffuseness are central claims, a representative sparse Bayesian comparison would strengthen the assessment of the learned prior’s value.

  2. Limited direct verification of reconstructed human activity. The real-data experiments validate localization using stimulation sites and resection cavities, not source amplitudes, temporal dynamics, or null-space components. The findings therefore support localization more strongly than faithful reconstruction of whole-brain activity. Moreover, improvements over strong baselines are not consistently statistically significant.

  3. The specific benefit of null-space learning is not fully isolated. Removing the null-space estimator demonstrates its usefulness relative to a model without that component, but does not establish that explicitly restricting learning to the null space is better than a comparably sized, anatomy-conditioned source estimator. The current comparisons do not fully separate the decomposition’s benefit from differences in architecture, denoising, or training.

  4. Simulated generalization is narrower than the broader reconstruction claims. The subject-wise split tests unseen anatomies, but training and testing use the same source-generation framework. It remains unclear how well the prior generalizes to unseen source-pattern families, different neural dynamics, or forward-model mismatch. This matters because the null-space estimate is determined by the learned prior rather than directly constrained by the measurements.

Questions for authors
  1. Were ConvDip and DeepSIF trained on the same simulated dataset with matched preprocessing and model selection? How were their outputs mapped to subject-specific source spaces?
  2. What evidence distinguishes the benefit of explicit null-space projection from that of a matched anatomy-conditioned network without the projection?
  3. How are the pseudo-inverse tolerance, EEG referencing, and signal normalization handled to maintain consistency between the measurements and lead-field matrix?
  4. Can the simultaneously recorded SEEG support any additional validation of reconstructed temporal activity?

Suggested Improvements

  1. please include a representative sparse Bayesian baseline such as Champagne or Gamma-MAP.
  2. please report Earth Mover’s Distance between normalized source-power maps on simulated data, specifying the spatial ground distance, to complement threshold-dependent metrics.
  3. please qualify claims of superiority where differences are not statistically significant, and distinguish unseen-anatomy generalization from unseen-dynamics generalization
4 comments

We sincerely thank the reviewer for the careful and constructive review. In response, we have added sparse Bayesian baselines, an ablation isolating the null-space restriction, generalization analyses under simulator mismatch, and validation beyond spatial localization on empirical data. All revisions are highlighted in blue in the revised manuscript; below, W, Q, and S denote Weaknesses, Questions, and Suggestions, respectively.


W1 & S1: Missing sparse Bayesian baselines

We have added γ-MAP [1] and Champagne [2], two representative sparse Bayesian ESI methods, as baselines across all three experimental paradigms: simulated source activity estimation (Simulation), intracranial stimulation localization (SPES), and epileptogenic zone estimation (EZ). CANDLE achieved better mean performance than both methods on all six metrics, with significant improvements in HD95 and Dice score in Simulation and in peak distance in SPES (p < 0.001). The complete results, including the remaining metrics and qualitative comparisons, are now reported in Figs. 3–5 and 11–13 and Tables 3–5 of the revised manuscript.

Exp.Metricγ-MAPChampagneCANDLE
SimulationHD95 [mm] ↓67.33 ± 0.77 ††^{\dagger\dagger}67.74 ± 0.77 ††^{\dagger\dagger}37.26 ± 0.74
SimulationDice score ↑0.12 ± 0.00 ††^{\dagger\dagger}0.12 ± 0.00 ††^{\dagger\dagger}0.48 ± 0.01
SPESSpatial dispersion [mm] ↓56.34 ± 1.7356.17 ± 1.6854.11 ± 1.25
SPESPeak distance [mm] ↓46.10 ± 2.43 ††^{\dagger\dagger}44.20 ± 2.26 ††^{\dagger\dagger}32.86 ± 1.91
EZHD95 [mm] ↓66.88 ± 3.9065.20 ± 4.4861.72 ± 2.75
EZCentroid distance [mm] ↓45.98 ± 3.9044.17 ± 4.2635.38 ± 2.98

W2 & Q4: Validation beyond localization on real data

Since the true whole-brain activity is unknown in empirical recordings, we focused on localizing brain regions directly targeted by experimenters or clinicians (e.g., the resection cavities). However, we agree that a more multifaceted validation is desirable. Of the three aspects you raised, we therefore assessed, albeit indirectly, (i) amplitude and (ii) temporal dynamics in the intracranial stimulation localization task:

  • Amplitude: Spearman correlation between the cortical shallowness of each SPES site and the peak strength of the ESI estimate (denoted as SPES site depth corr.). As ESI estimates are unitless, this serves as a proxy, assuming that estimated strength correlates with the underlying potential.
  • Temporal dynamics: normalized MAE/MSE between the SEEG waveform at the contact nearest to the SPES electrode and the estimated waveform in the nearest parcel (SEEG MAE/MSE).

The remaining aspect, (iii) the null-space component cannot be validated directly on empirical recording, as this requires the true whole-brain activity. However, its effectiveness is supported by Models (b-ii) and (c-ii) in Figure 6. As shown in the table below, CANDLE outperformed or was comparable to existing methods across all three metrics. These results suggest that, beyond spatial localization, CANDLE is also useful for a more multifaceted assessment of brain activity related to amplitude and temporal dynamics. We added these results to Appendix D.3 and Table 2 of the revised manuscript.

MetricLCMVsLORETAwMNEγ-MAPChampagneeLORETAConvDipDeepSIFGBFCANDLE
SPES site depth corr. ↑0.10±0.100.10 \pm 0.100.16±0.090.16 \pm 0.090.17±0.100.17 \pm 0.100.19±0.100.19 \pm 0.100.19±0.100.19 \pm 0.100.12±0.140.12 \pm 0.140.20±0.090.20 \pm 0.090.18±0.090.18 \pm 0.090.14±0.080.14 \pm 0.080.23±0.09\mathbf{0.23 \pm 0.09}
SEEG MAE ↓0.40±0.00††0.40 \pm 0.00^{\dagger\dagger}0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.34±0.01††0.34 \pm 0.01^{\dagger\dagger}0.34±0.01††0.34 \pm 0.01^{\dagger\dagger}0.34±0.01††0.34 \pm 0.01^{\dagger\dagger}0.40±0.00††0.40 \pm 0.00^{\dagger\dagger}0.37±0.00††0.37 \pm 0.00^{\dagger\dagger}0.32±0.01\mathbf{0.32 \pm 0.01}0.32±0.01\mathbf{0.32 \pm 0.01}
SEEG MSE ↓0.25±0.01††0.25 \pm 0.01^{\dagger\dagger}0.18±0.01††0.18 \pm 0.01^{\dagger\dagger}0.19±0.01††0.19 \pm 0.01^{\dagger\dagger}0.19±0.01††0.19 \pm 0.01^{\dagger\dagger}0.19±0.01††0.19 \pm 0.01^{\dagger\dagger}0.19±0.01††0.19 \pm 0.01^{\dagger\dagger}0.25±0.00††0.25 \pm 0.00^{\dagger\dagger}0.22±0.01††0.22 \pm 0.01^{\dagger\dagger}0.18±0.01††0.18 \pm 0.01^{\dagger\dagger}0.17±0.00\mathbf{0.17 \pm 0.00}

W3 & Q2: Isolating the benefit of the null-space restriction

To disentangle the benefit of restricting learning to the null space from that of conditioning on subject-specific cortical geometry in CANDLE, we conducted an additional ablation, which we added to Subsection 3.4 of the revised manuscript as Model (b-iii). Please refer to our response to Reviewer PYox (W1-(i)) for details.

W4: Scope of generalization in simulations

In the simulated source activity estimation, we have explicitly stated that the scope of CANDLE's generalization is to unseen anatomies (L334). Nevertheless, we agree that evaluating this scope alone is insufficient, and we additionally evaluated generalization along the three axes you raised under the following conditions:

  • Unseen source configurations: cluster-wise split by source configuration.
  • Unseen neural dynamics: test source activity generated with the Wilson–Cowan[3] or Stuart–Landau[4] model instead of the original Jansen–Rit model [5].
  • Forward-model mismatch: test EEG synthesized with electrode jitter or montage shift of 5, 10, and 15 mm.

As summarized in the table below, CANDLE achieved the best Dice score under all conditions, with statistically significant improvements over all baselines (p<0.001p < 0.001). Although the performance of CANDLE degraded gradually with the degree of shift, its advantage over the baselines was maintained, suggesting that the learned null-space prior remains effective beyond the source-generation framework used for training. Results for all baselines and metrics, detailed settings, and further discussion are provided in Appendix D.1 and Figure 9 of the revised manuscript.

ConditionConvDipDeepSIFCANDLE
Unseen anatomies (Subsection 3.1)0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.48±0.01\mathbf{0.48 \pm 0.01}
Unseen source configurations0.31±0.00††0.31 \pm 0.00^{\dagger\dagger}0.31±0.01††0.31 \pm 0.01^{\dagger\dagger}0.44±0.01\mathbf{0.44 \pm 0.01}
Unseen neural dynamics (Wilson–Cowan model)0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.33±0.01††0.33 \pm 0.01^{\dagger\dagger}0.48±0.01\mathbf{0.48 \pm 0.01}
Unseen neural dynamics (Stuart–Landau model)0.25±0.00††0.25 \pm 0.00^{\dagger\dagger}0.27±0.01††0.27 \pm 0.01^{\dagger\dagger}0.31±0.00\mathbf{0.31 \pm 0.00}
Forward-model mismatch (Electrode jitter, 5 mm)0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.48±0.01\mathbf{0.48 \pm 0.01}
Forward-model mismatch (Electrode jitter, 15 mm)0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.46±0.01\mathbf{0.46 \pm 0.01}
Forward-model mismatch (Montage shift, 5 mm)0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.32±0.01††0.32 \pm 0.01^{\dagger\dagger}0.47±0.01\mathbf{0.47 \pm 0.01}
Forward-model mismatch (Montage shift, 15 mm)0.31±0.01††0.31 \pm 0.01^{\dagger\dagger}0.31±0.01††0.31 \pm 0.01^{\dagger\dagger}0.43±0.01\mathbf{0.43 \pm 0.01}

Q1: Implementation details of ConvDip / DeepSIF baselines

ConvDip and DeepSIF were trained on simulated data under the same settings as CANDLE and then evaluated zero-shot on empirical data. Please refer to our response to Reviewer PYox (W1-(iii)) for details. As this was not explicitly stated in the original manuscript, we have added a clarification to Appendix C.1 of the revised manuscript.


Q3: Consistency between EEG measurements and the lead-field matrix

For the pseudoinverse of the lead-field matrix L\mathbf{L}, we use the Moore–Penrose pseudoinverse with a tolerance of 1.2×10−71.2 \times 10^{-7} (the epsilon of PyTorch’s float32 type), chosen for numerical stability. For EEG referencing, we apply common average referencing (CAR) to the EEG, as described in Appendix B.2. We apply the same operator to the lead-field matrix, L′=(I−11⊤/C)L\mathbf{L}' = (\mathbf{I} − \mathbf{11}^\top/C)\mathbf{L}, so the measurements and the forward model share the same reference. For signal normalization, we apply z-score normalization to the EEG only, using a single mean and standard deviation computed across all channels and time points (L′\mathbf{L}' is left unnormalized). Because this normalization is applied after CAR, it changes only the overall amplitude scale and otherwise preserves consistency with L′\mathbf{L}'. As discussed in our response to W2, the source activity estimated by CANDLE, like that of the learning-based ESI baselines, is a dimensionless feature rather than a physical quantity. The absolute scale therefore does not affect the source localization tasks evaluated in this work. As these implementation details were not clearly stated in the original manuscript, we have added them to Appendix B.2 of the revised version.

S2: Earth Mover’s Distance on simulated data

As a threshold-independent evaluation of spatial localization, the table below reports the Earth Mover’s Distance (EMD) on simulated data. Source distributions were obtained by normalizing the temporal mean of the squared source activity in each region, and the ground distance was defined as the Euclidean distance between region centroids in each subject’s cortical space. CANDLE achieved an EMD of 27.55 mm, significantly lower than all baselines (p<0.001p < 0.001). This suggests that CANDLE’s advantage does not depend on thresholding and that the entire estimated distribution is concentrated in the vicinity of the true sources. We added these results to Appendix D.2 and Figure 10 of the revised manuscript.

LCMVsLORETAwMNEγ-MAPChampagneeLORETAConvDipDeepSIFGBFCANDLE
EMD [mm]46.60±0.43††46.60 \pm 0.43^{\dagger\dagger}36.31±0.30††36.31 \pm 0.30^{\dagger\dagger}38.46±0.30††38.46 \pm 0.30^{\dagger\dagger}32.76±0.35††32.76 \pm 0.35^{\dagger\dagger}32.82±0.36††32.82 \pm 0.36^{\dagger\dagger}40.12±0.29††40.12 \pm 0.29^{\dagger\dagger}40.00±0.47††40.00 \pm 0.47^{\dagger\dagger}41.46±0.42††41.46 \pm 0.42^{\dagger\dagger}34.30±0.33††34.30 \pm 0.33^{\dagger\dagger}27.55±0.41\mathbf{27.55} \pm \mathbf{0.41}

S3: Qualifying superiority claims and distinguishing types of generalization

We have qualified claims of superiority where differences were not statistically significant (L370–372, L453–454). Regarding the distinction between unseen-anatomy and unseen-dynamics generalization, for simulated data we explicitly state that our experiment evaluates generalization to unseen anatomies (L334). For empirical data, both anatomy and dynamics are unseen, so the generalization cannot be attributed to either factor alone. We agree that generalization to unseen dynamics was insufficiently evaluated, and we have added such an evaluation (please refer to our response to W4 and Appendix D.1 of the revised manuscript).


References

[1] David Wipf and Srikantan Nagarajan. A unified Bayesian framework for MEG/EEG source imaging. NeuroImage, 44(3):947–966, 2009.

[2] David P Wipf, Julia P Owen, Hagai T Attias, Kensuke Sekihara, and Srikantan S Nagarajan. Robust Bayesian estimation of the location, orientation, and time course of multiple correlated neural sources using MEG. NeuroImage, 49(1):641–655, 2010.

[3] Hugh R Wilson and Jack D Cowan. Excitatory and inhibitory interactions in localized populations of model neurons. Biophys. J., 12(1):1–24, 1972.

[4] Anton A Selivanov, Judith Lehnert, Thomas Dahms, Philipp H ̈ovel, Alexander L Fradkov, and Eckehard Sch ̈oll. Adaptive synchronization in delay-coupled networks of Stuart-Landau oscillators. Phys. Rev. E., 85(1):016201, 2012.

[5] Ben H Jansen and Vincent G Rit. Electroencephalogram and visual evoked potential generation in a mathematical model of coupled cortical columns. Biol. Cybern., 73(4):357–366, 1995.

As the rebuttal period is drawing to a close, we would appreciate hearing whether our responses have adequately addressed your concerns. If any issues remain, we would be happy to discuss them further. Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Rebuttal

General Response: Summary of Revisions

We sincerely thank the Area Chair for overseeing the review process and the reviewers for their constructive feedback. Reviewers sBb5 and xRnr initially recommended Weak Acceptance. Following our rebuttal, Reviewer xRnr acknowledged that our responses addressed most of their questions and reaffirmed their recommendation for acceptance. Reviewer PYox also explicitly confirmed that "there are no remaining issues" after reviewing our responses.

During the rebuttal period, we conducted extensive additional experiments and revised the manuscript to address the reviewer's concerns. All changes are highlighted in blue in the revised manuscript. Below, we summarize the key revisions.

1. Isolating the Null-Space Constraint (Reviewers sBb5, xRnr, PYox)

  • Controlled Ablation: We introduced a new ablation, Model (b-iii), which removes the null-space projection while preserving the same network architecture, parameter count, and training configuration. This directly isolates the contribution of restricting learning to the null space. CANDLE outperformed this unrestricted variant across all four metrics on two empirical tasks. These results have been incorporated into Section 3.4 of the revised manuscript.

2. Strengthening Baseline Evaluation (Reviewers sBb5, PYox)

  • Additional Bayesian Baselines: We added γ\gamma-MAP and Champagne, two representative sparse Bayesian ESI methods, across all three experimental tasks. CANDLE achieved better mean performance on all six metrics, with statistically significant improvements in simulated localization and SPES peak distance. These results are reported in Figures 3–5 and 11–13 and Tables 3–5 of the revised manuscript.
  • Baseline Implementation Details: We clarified that ConvDip and DeepSIF used the same simulated dataset, preprocessing, data splits, and model selection criteria as CANDLE. We also specified how their predictions are mapped to subject-specific cortical source spaces through atlas label correspondence. These details are documented in Appendix C.1 of the revised manuscript.

3. Temporal Validation and Generalization (Reviewers sBb5, xRnr)

  • Temporal Dynamics Validation: We introduced two complementary evaluations of temporal dynamics: frequency-specific functional connectivity (FC) correlation on simulated data and normalized waveform MAE/MSE against simultaneously recorded SEEG on empirical data. CANDLE achieved the highest FC correlation across all four evaluated frequency ranges, significantly outperforming all baselines in the broadband and δ\delta bands (p<0.01p < 0.01), and the lowest SEEG MSE. These results provide additional evidence beyond spatial localization and are reported in Section 3.1, Figure 3, and Appendix C.2 (simulation) and Appendix D.3 and Table 2 (empirical data) of the revised manuscript.
  • Generalization Under Distribution Shifts: We evaluated generalization to unseen source configurations, alternative neural dynamics (Wilson–Cowan and Stuart–Landau), and forward-model mismatch caused by electrode jitter and montage shifts. CANDLE consistently achieved the highest Dice scores across all conditions, significantly outperforming the baselines (p<0.001p < 0.001). These results are reported in Appendix D.1 and Figure 9 of the revised manuscript.
  • Robustness to EEG Channel Reduction: We evaluated CANDLE's robustness to reduced EEG channel counts (64, 32, and 16 channels) on simulated and empirical data. CANDLE consistently outperformed the baselines across all channel counts, demonstrating its robustness to reduced sensor density. These results are reported in Section 3.5 of the revised manuscript.

We sincerely appreciate the time and consideration of the Area Chair and all reviewers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.