{"id":"4854a1a9-4b0f-4a39-a9d2-9041ba6d365e","arxiv_id":"1908.04118","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D U-net can reconstruct brain conductivity maps from MRI phase data, but it only generalizes when training anatomy and artifacts match the target data.","lead":"This paper tests whether a 3D deep learning network can map brain tissue conductivity from MRI phase data, and finds that networks trained on computer simulations produce artifacts on real patient scans. Training on real scans with conventional conductivity labels removes many artifacts, but only for data that closely resembles the training set.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-vivo evaluation is partly circular: conductivity labels and the evaluation reference are both the same HHEPT estimator (Eq. 1), so the reported artifact reduction could reflect the network learning to reproduce HHEPT rather than recovering true conductivity.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: in-vivo training labels and the evaluation reference are both HHEPT, so improved correlation with HHEPT does not establish improved accuracy of conductivity recovery. This is the single most load-bearing issue because the paper's conclusion about artifact reduction in vivo depends on the evaluation metric being a valid proxy for true conductivity. If HHEPT is systematically biased, then both the training signal and the scoring penalize deviation from that bias, and the reported 'artifact reduction' may be nothing more than alignment with the HHEPT reconstruction pipeline. The paper itself flags the limitation in the Discussion, which supports the seriousness of the concern rather than resolving it. The simulation experiments with known ground truth are genuine independent evidence that DLEPT can learn the Helmholtz inversion from clean simulated data, but they do not validate the in-vivo generalization claim. Other issues, such as missing hyperparameters and lack of released code or data, are secondary to this circularity. The reader's CONDITIONAL verdict already captures this risk, so no change in verdict is needed; the concern confirms rather than alters the assessment.","tokens_in":7582,"tokens_out":2894,"duration_ms":32996,"concrete_test":"Retain the same three-fold split, but replace the in-vivo evaluation reference with an independent conductivity estimate for the same subjects: e.g., tissue segmentation into gray matter, white matter, and CSF with literature conductivity values, or a non-HHEPT reconstruction such as contrast-source-inversion EPT. Recompute Table 2 correlations of NWvol, NWpat, and NWvol+pat against this independent reference and compare with the HHEPT correlation. If the improvement over simulation-trained NWDE persists against the independent reference, the artifact-reduction claim is about true conductivity; if only HHEPT correlation improves, the central in-vivo claim is largely circular.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central in-vivo claim is that training on realistic phase data with 'conductivity labels from conventional EPT' severely reduces reconstruction artifacts. But the labels and the quantitative evaluation reference are both HHEPT maps computed from Eq. (1) with the transceive-phase assumption and the median-filtered method of reference [4]. Therefore the evaluation measures agreement with the very estimator used to create training targets. A network trained to regress HHEPT labels will, by construction, produce outputs that correlate with HHEPT; the reported improvement may be mimicry of the HHEPT prior, including its smoothing and boundary errors, rather than robustness to acquisition artifacts. The discussion acknowledges this limitation: 'This is unfortunately not guaranteed for in-vivo reconstructions using networks trained with HHEPT, given the intrinsic inaccuracies of this technique.' The simulation experiments provide independent support for DLEPT's ability to recover known conductivity, but they do not break the circularity of the in-vivo evaluation. In particular, the CSF artifacts that motivate the study may themselves be present in the HHEPT references, since CSF pulsation affects the acquired phase [26], so reducing deviation from HHEPT may not reduce true error. The claim as stated—'training with realistic phase data and conductivity labels from conventional EPT allows for severely reducing these artifacts'—is therefore not independently established for in-vivo data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates a 3D patch-based U-net for brain conductivity mapping from B1 transceive phase data (DLEPT). The authors train networks on electromagnetic simulations (Duke/Ella models) with and without added Gaussian noise and test on simulated and in-vivo data (healthy volunteers and patients with brain lesions). They also train networks on in-vivo data using conductivity labels from conventional Helmholtz-based EPT (HHEPT) and evaluate by correlation with HHEPT references. The main findings are: (1) networks trained on simulations reconstruct simulated data well but produce artifacts when applied to in-vivo data, especially in CSF; (2) training on in-vivo data with HHEPT labels reduces these artifacts; and (3) generalization between volunteer and patient datasets is limited, suggesting sensitivity to geometry and acquisition-related artifacts.","tokens_in":7882,"tokens_out":2564,"duration_ms":27587,"significance":"If the central claims hold, the paper would be a useful contribution to EPT by quantifying the domain-shift problem in deep-learning-based conductivity mapping and by showing that training on in-vivo-like data improves robustness to acquisition artifacts. The study has several strengths: it uses realistic electromagnetic simulations with known ground truth, it applies cross-validation and excludes validation data from training, it includes pathology, and it explicitly discusses the limitations of HHEPT labels. The negative result that sim-trained networks do not transfer to in-vivo data is valuable for the community. However, the in-vivo evaluation is partly circular because the same HHEPT estimator provides both the training labels and the evaluation reference, so the reported artifact reduction is not an independent measure of true conductivity accuracy. This limits the support for the paper's central conclusion.","major_comments":[{"comment":"The in-vivo networks are trained on conductivity labels from HHEPT, computed from Eq. (1) with the transceive-phase assumption and median-filtered method of reference [4], and are then evaluated by correlation with the same HHEPT reference. This makes the reported artifact reduction partly a measure of agreement with the training-target estimator. Since HHEPT has known systematic errors at tissue boundaries and is affected by CSF pulsation (as the authors note in the Discussion and reference [26]), the central claim that \"training with realistic phase data and conductivity labels from conventional EPT allows for severely reducing these artifacts\" is not independently established. The Discussion does acknowledge this limitation, but the Abstract and Conclusion still present the artifact reduction as a robust finding. To support the claim, the authors should either (a) evaluate on simulated data that mimic in-vivo phase artifacts and have known ground-truth conductivity, (b) use an independent reference (e.g., tissue-segmentation-based literature conductivity values or ex-vivo measurements), or (c) explicitly reframe the conclusion as \"reduces deviation from the HHEPT estimator\" rather than \"reduces artifacts\" in an absolute sense.","section":"Methods/Experimental outline and Results (Tab. 2, Fig. 4, Fig. 5)"},{"comment":"The quantitative evaluation reports only average correlation coefficients without confidence intervals, per-fold ranges, or statistical significance tests. With 14 patients and 18 volunteers, the observed differences between NWvol, NWpat, and NWvol+pat could be within sampling variability. The authors should provide per-subject correlations, standard deviations or bootstrap confidence intervals, and, where appropriate, paired significance tests. This is particularly important for the cross-validation comparisons in Tab. 3, where the claim that accuracy saturates with half the training data rests on average correlations alone.","section":"Results (Tab. 2, Tab. 3)"}],"minor_comments":[{"comment":"The abstract mentions \"different levels of homogeneous Gaussian noise introduced in training and testing,\" but the methods describe only specific SNR values (100 and 200). Please clarify which SNR values were used for each network and why the level was changed from 100 to 200.","section":"Abstract and Methods/Experimental outline"},{"comment":"The network architecture figure would be easier to interpret with a legend or caption listing the patch size, number of channels per layer, and the locations of residual units and batch normalization, as these details are described only in the text.","section":"Fig. 1"},{"comment":"The notation Δφ+ is introduced without a formal definition of the Laplacian convention or the transceive-phase assumption. Adding a sentence defining these terms would improve reproducibility.","section":"Theory, Eq. (1)"},{"comment":"Tab. 1 lists the lesions included in the patient dataset, but the table content is not described in the text. Please refer to it explicitly and state how lesion presence might affect the reconstruction evaluation.","section":"Results, Tab. 1"},{"comment":"The comparison between DSpat and DSvol uses sdR as a proxy for brain-shape differences, but the volunteer and patient data were acquired at different sites and likely with different populations. The observed generalization gap could be confounded by acquisition-site or sequence differences. A sentence acknowledging this confound and suggesting a matched-site study would strengthen the interpretation.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The circularity in the in-vivo evaluation is the main substantive issue. The authors are honest about it in the Discussion, but the framing of the Abstract and Conclusion overstates what is demonstrated. I would be willing to see a revised version that either provides an independent form of in-vivo validation or clearly restricts the claims to agreement with HHEPT. The simulation results and the careful cross-validation design are valuable and should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a genuine step forward for the EPT-deep-learning subfield, though it is an extension rather than a new framework. The 3D patch-based U-net, phase-only input, and the systematic cross-dataset generalization study (sim, healthy volunteer, patient, with pathologies) are all useful. The simulation experiments are the strongest part: ground truth is available, and the results show the expected degradation with geometric mismatch and with added noise. The negative result that homogeneous Gaussian noise in training does not transfer to in-vivo artifacts is valuable and clearly presented.\n\nThe main soft spot is the circularity in the in-vivo evaluation. The networks are trained on HHEPT conductivity labels, and then scored by correlation with HHEPT maps computed from the same formula (Eq. 1) using the same median-filtered method. So part of the reported artifact reduction is the network learning to reproduce HHEPT—including its smoothing and boundary errors—rather than a demonstrated recovery of true conductivity. The paper does acknowledge this in the discussion: 'This is unfortunately not guaranteed for in-vivo reconstructions using networks trained with HHEPT, given the intrinsic inaccuracies of this technique.' But the abstract and conclusion state the artifact-reduction claim more strongly than the evidence supports. The CSF artifacts that motivate the study may be present in the HHEPT references themselves, so reducing deviation from HHEPT does not necessarily reduce true error. This is a real limitation, but not a disqualifying one, because the simulation experiments provide independent evidence that DLEPT can recover known conductivity.\n\nOther issues are minor but worth noting: no code or data are provided, key training hyperparameters are only partially reported, and there are no confidence intervals on the correlation values. These are fixable in revision.\n\nWho is this for? Researchers working on EPT, deep-learning reconstruction, or phase-based conductivity mapping. It is a well-scoped empirical study with an honest discussion and a clear message: simulation-trained networks do not generalize to in-vivo artifacts, and realistic training data helps. The central limitation is not hidden, and the findings are plausible. It deserves a serious referee. I would recommend sending it to peer review, with the expectation that the in-vivo claims be softened or supported by additional validation (e.g., phantom data, synthetic in-vivo-like artifacts, or a clear statement that the in-vivo evaluation is only against the training target).","headline":"Solid extension of DLEPT with honest limits: the simulation story is clean, but the in-vivo evaluation leans on HHEPT for both labels and reference, so the central artifact-reduction claim is only partly established.","tokens_in":8381,"tokens_out":1338,"would_cite":true,"duration_ms":16472,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Simulation-trained networks fail on real brain scans","keywords":["deep learning","electrical properties tomography","brain conductivity","3D U-net","patch-based reconstruction","transceive phase","in-vivo generalization","MRI"],"falsifier":"A concrete check is to run the same simulation-trained network on a physical phantom or numerical phantom with known ground-truth conductivity and realistic in-vivo-like phase artifacts (motion, CSF pulsation). If artifacts still appear despite matching the phantom's anatomy and artifacts, the claimed role of in-vivo training data would be weakened; conversely, simulating those artifacts and seeing artifacts disappear would confirm it.","tokens_in":7406,"feed_emoji":"🧠","tokens_out":7619,"duration_ms":70425,"temperature":0.7,"pith_summary":"This paper studies how far a deep-learning network can go in turning MRI radio-frequency phase maps into brain conductivity maps, a technique called deep-learning electrical properties tomography. It reports that a 3D patch-based U-net trained on electromagnetic simulations reconstructs simulated brains with high fidelity, but when the same network is applied to real volunteer and patient scans it produces visible artifacts, especially in cerebrospinal fluid. Adding homogeneous Gaussian noise to simulated training data does not fix the problem. The paper then shows that training the same architecture on real in-vivo phase maps, using conventional Helmholtz-based EPT reconstructions as conductivity labels, sharply reduces those artifacts. If the authors are right, deep-learning EPT is clinically promising but only when the training data closely matches the target anatomy, acquisition noise, and tissue types.","feed_headline":"Simulation-trained networks fail on real brain scans","feed_subtitle":"Training on real patient phase data with standard conductivity labels cuts artifacts, but only for matched anatomy.","key_machinery":"The load-bearing machinery is two-part. The first part is the Helmholtz-based phase-only conductivity reconstruction, HHEPT, which computes conductivity as $\\sigma = \\Delta\\phi^{+} / (\\mu_0 \\omega)$ from the Laplacian of the B1 transceive phase under the transceive-phase assumption; the paper uses HHEPT maps as both the in-vivo training labels and the evaluation reference. The second part is a 3D patch-based U-net with two downsampling steps, three convolutions between each step, residual units, batch normalization, and ReLU activations, trained on 24x24x24 phase patches with mean subtraction and overlapping-patch averaging. The network learns a surrogate for the inverse mapping from local B1 phase patches to conductivity, and the key experimental lever is which data supplies its training patches: pure simulations with added Gaussian noise, or real in-vivo phase maps carrying genuine acquisition artifacts.","core_discovery":"The central claim is that the ability of DLEPT networks to generalize is governed less by network architecture than by how faithfully the training data reproduces the measurement conditions of the target data. Simulation-trained networks achieve correlations up to 0.991 on simulated test data, but applying them to in-vivo data produces artifact-heavy reconstructions, and training them with added Gaussian noise (even at the empirically optimal SNR=200) degrades generalization to geometries not present in training. Networks trained on in-vivo phase data, with conductivity labels provided by HHEPT reconstructions, reconstruct both volunteers and patients with far fewer artifacts and higher correlation with the HHEPT reference. The authors conclude that in-vivo training works because the training phase maps already contain acquisition-related artifacts such as head motion and CSF pulsation, and that simulation-trained networks will need realistic simulated artifacts, not just homogeneous noise, to become clinically transferable.","pith_inferences":["If HHEPT labels carry boundary or pulsation errors, then the reported artifact reduction may partly be the network learning to reproduce HHEPT's own artifacts; replacing labels with more accurate references (segmented literature values, ex-vivo measurements, or forward-model fits) would tell how much of the gain is real conductivity recovery.","The stronger dependence on anatomy and artifacts suggests a practical deployment path of fine-tuning or domain adaptation on local scanner data rather than relying on one universal model; this follows from the cross-population performance drop but is not tested in the paper.","A testable extension is to simulate CSF pulsation and head motion as phase perturbations in the synthetic training data; if the artifact reduction seen with in-vivo training can be reproduced synthetically, simulation-based DLEPT could be made transferable without acquiring large labeled patient cohorts."],"forward_implications":["Performance on simulated test data does not predict performance on real scans: correlations above 0.99 drop to visibly artifact-laden reconstructions when the same network is applied in vivo.","Training on in-vivo data, with HHEPT maps as labels, suppresses acquisition-related artifacts in the network output because the training phase maps already contain those artifacts.","Networks trained on one in-vivo population (volunteers) do not generalize to another (patients), and combining both populations gives the best results, so training sets must span the intended target group.","Adding homogeneous Gaussian noise during simulation training does not bridge the simulation-to-clinic gap; realistic artifact simulation is the missing ingredient."],"supporting_citations":[{"why":"Supplies the median-filtered HHEPT reconstruction used to create in-vivo conductivity labels and the evaluation reference.","marker":"[4]"},{"why":"Prior DLEPT feasibility study whose simulation setup and U-net approach this work extends to in-vivo generalization.","marker":"[18]"},{"why":"Dictionary-based machine-learning EPT approach that motivates local patch-based learning from B1 phase patterns.","marker":"[19]"},{"why":"Provides the anatomical models used for the electromagnetic simulations that make up the simulated training set.","marker":"[24]"},{"why":"Specifies the bSSFP sequence used to acquire the in-vivo transceive phase data.","marker":"[25]"},{"why":"Documents the effect of CSF pulsation on reconstructed brain conductivity, which the paper invokes to explain in-vivo artifacts.","marker":"[26]"},{"why":"Defines the Helmholtz EPT model and transceive-phase assumption underlying the HHEPT reference.","marker":"[1]"}],"fun_headline_variants":["Sim-only training fails on real brain scans","Real data beats simulation for brain mapping","Simulation-trained DLEPT artifacts on in-vivo","In-vivo training reduces EPT artifacts","Training data choice key for brain conductivity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the HHEPT conductivity maps used as in-vivo training labels and as the evaluation reference are accurate enough to stand in for true tissue conductivity; if HHEPT is systematically wrong at tissue boundaries or under CSF pulsation, the network learns and is scored by those same errors.","fun_headline_variants_meta":{"raw":{"variants":["Sim-only training fails on real brain scans","Real data beats simulation for brain mapping","Simulation-trained DLEPT artifacts on in-vivo","In-vivo training reduces EPT artifacts","Training data choice key for brain conductivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1337,"prompt_tokens":964,"completion_tokens":373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":580,"tokens_out":373,"duration_ms":4598,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:50:41.823149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check is to run the same simulation-trained network on a physical phantom or numerical phantom with known ground-truth conductivity and realistic in-vivo-like phase artifacts (motion, CSF pulsation). If artifacts still appear despite matching the phantom's anatomy and artifacts, the claimed role of in-vivo training data would be weakened; conversely, simulating those artifacts and seeing artifacts disappear would confirm it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the median-filtered HHEPT reconstruction used to create in-vivo conductivity labels and the evaluation reference."},{"cited_title":"Opening a new window on MR-based Electrical Properties Tomography with deep learning","cited_arxiv_id":"1804.00016","evidence_quote":"Prior DLEPT feasibility study whose simulation setup and U-net approach this work extends to in-vivo generalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Helmholtz EPT model and transceive-phase assumption underlying the HHEPT reference."}],"review_version":1}