{"id":"463d4530-36c7-4364-9c98-5897aec20b23","arxiv_id":"2505.22685","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DeepMultiConnectome classifies tractography streamlines by the brain regions they connect, producing 84- and 164-region connectomes in about 40 seconds per scan, with correlation r=0.992 and r=0.986 to traditionally generated connectomes.","lead":"A deep-learning model predicts brain connectivity maps directly from fiber-tracking data, skipping the slow brain-parcellation step. It generates two standard connectome atlases from a whole scan in about 40 seconds, with maps similar to conventional ones (correlation about 0.99).","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline r=0.992/0.986 may be trivially high: a group-average connectome baseline could match or nearly match these values, so subject-specific predictive value is unproven.","rationale":"The reader's weakest_assumption focused on whether 15-point streamline geometry can resolve ROI pairs and whether tck2connectome labels are the right ceiling. Those are real issues, but they are downstream of a more fundamental question: does the reported similarity metric actually measure subject-specific predictive accuracy? The reader did note that the correlation may be inflated by high-count edges given the low macro F1, but did not identify the more decisive check: comparing against a group-average connectome baseline. My concern extends the reader's intuition into a concrete, falsifiable test that directly targets the strongest claim. If the group-average baseline matches DeepMultiConnectome's correlation, then the central claim of a 'fast surrogate for the conventional pipeline' is not supported, regardless of per-streamline classification quality. This is not an ad hominem or a disagreement with consensus; it is a question of whether the evaluation design can distinguish the model from a much simpler null model. The paper itself flags insufficient sensitivity of evaluation metrics in the Discussion, which is in-scope evidence supporting the concern. The verdict remains CONDITIONAL because the concern is testable and not yet resolved; no change to the reader's verdict is needed, but the condition should explicitly include this baseline comparison.","tokens_in":14216,"tokens_out":3970,"duration_ms":42331,"concrete_test":"Recompute Section 3.2 similarity using a baseline that assigns every test subject the group-average traditional connectome (mean of training-subject connectomes, same parcellation and preprocessing). Report Pearson r and LERM for this baseline versus each subject's traditional connectome, using the same upper-triangular and SPD handling as the paper. Additionally, recompute the reported r restricted to edges present in at least 50% of subjects, and on log-transformed edge counts. If the baseline r is within 0.01 of DeepMultiConnectome's r (0.992/0.986), the headline correlation does not demonstrate subject-specific prediction; if the baseline is substantially lower, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 3.2 is that predicted connectomes are 'highly correlated' with traditional ones (r=0.992±0.003 for 84 ROIs, r=0.986±0.004 for 164 ROIs), and the paper interprets this as evidence that DeepMultiConnectome can replace the parcellation-based pipeline. The load-bearing weakness is that this metric may not distinguish the model from a trivial baseline. Structural connectomes are highly stereotyped across healthy young adults; the paper's own Figure 4 shows intrasubject similarity is higher than intersubject similarity, but it never reports the actual intersubject correlation values or compares against the simplest possible predictor: the group-average (mean training) connectome assigned to every subject. With sparse connectome matrices, Pearson correlation over upper-triangular elements is dominated by zero-zero pairs and a few high-count edges. The low macro F1 scores (50.48% for 84 ROIs, 34.36% for 164 ROIs) show that many streamlines are misclassified; if errors concentrate in low-count or absent edges, the connectome-level r remains almost unchanged. The paper's Discussion explicitly concedes that 'our evaluation methods may not be sufficiently sensitive, potentially overestimating the true quality of the predictions.' That self-flagged limitation is exactly the risk here: without a group-average baseline and without edge-wise or sparsity-aware metrics, the headline r does not establish that the model adds subject-specific information beyond the population mean. If a template connectome yields r close to 0.99, the speed advantage reduces to 'output the average matrix,' and the claimed practical value as a subject-specific surrogate collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepMultiConnectome, a point-cloud neural network that classifies tractography streamlines directly into gray-matter region pairs for two parcellation schemes (84 and 164 ROIs) simultaneously, thereby bypassing anatomical parcellation. Trained on 1,000 HCP-YA subjects with labels derived from MRtrix's tck2connectome, the model predicts connectomes from approximately 2.93 million streamlines per subject in about 40 seconds. On 200 held-out test subjects, predicted connectomes achieve Pearson correlations of r=0.992 (84 ROI) and r=0.986 (164 ROI) with traditional connectomes, largely preserve six network measures, show test-retest reproducibility statistically indistinguishable from the traditional pipeline, and yield comparable age/cognition prediction performance in a downstream 1D-CNN. The authors conclude that DeepMultiConnectome provides a scalable and fast surrogate for the conventional parcellation-and-labeling connectome pipeline.","tokens_in":14428,"tokens_out":4937,"duration_ms":51390,"significance":"If the subject-specific predictive value is firmly established, the method is valuable: it reduces per-subject connectome generation from hours to about 40 seconds, removes the FreeSurfer parcellation bottleneck, and supports multiple parcellation schemes in a single model. The evaluation is out-of-sample on 200 unseen subjects and includes several complementary checks—intrasubject versus intersubject similarity, network-measure correlations, test-retest parity, and a downstream phenotype prediction experiment—which are appropriate and largely behave as expected for a successful surrogate. The paper is also transparent about its limitations, including the explicit concession in the Discussion that its evaluation methods may overestimate true prediction quality. However, the central quantitative claim currently rests on Pearson correlations over sparse connectome matrices, which can be inflated by stereotypy and zero-zero agreement; the absence of a group-average baseline and edge-wise analyses leaves the subject-specificity of the predictions not fully established.","major_comments":[{"comment":"The headline r=0.992/0.986 is never compared against the simplest null model: assigning the mean training connectome to every test subject. Because structural connectomes are highly stereotyped across healthy young adults and the matrices are sparse, Pearson correlation over upper-triangular elements can be high even for a group-average predictor. The paper reports intrasubject versus intersubject similarity only as distributions in Figure 4 without giving the actual intersubject correlation values. Please report those values and add a group-average baseline to the same evaluation. This is load-bearing for the claim that the predictions are subject-specific rather than reflections of population-level connectivity.","section":"Section 3.2, Figure 4, Table 3"},{"comment":"The macro F1 scores of 50.48% (84 ROI) and 34.36% (164 ROI) indicate that the majority of streamline classes are misclassified at the per-class level. The paper hypothesizes that these errors are localized to low-impact edges or that the evaluation is insensitive, but this is not demonstrated. Please provide edge-wise analyses: for example, prediction error stratified by true edge count, Pearson correlation computed only over nonzero edges, and confusion statistics for low-count and long-range edges. Without such analyses, the high connectome-level r could be carried by a few high-count edges and by agreement on zero-zero pairs, and the stated macro F1 makes this a concrete risk.","section":"Section 3.1, Tables 1 and 2; Section 3.2"},{"comment":"The test-retest parity between DeepMultiConnectome and the traditional method is presented as evidence of reproducibility, but both methods are evaluated against the same tck2connectome labeling procedure, and a group-average predictor would also show high test-retest correlation because the same group mean would be assigned to both sessions. Please add a group-average baseline to Table 4 and report the actual intersubject test-retest or between-subject similarity values. This would allow the parity to be interpreted as subject-specific reliability rather than as a consequence of stereotypy.","section":"Section 3.3, Table 4"},{"comment":"The training labels and the evaluation target are both generated by tck2connectome's nearest-endpoint assignment from the same parcellation. The paper should explicitly state that the reported r measures agreement with this specific automated pipeline and not with an independent anatomical ground truth, and it should discuss how this ceiling affects the claim that the model can replace the traditional pipeline. The Discussion partially acknowledges a ceiling, but the statement belongs in the methods and evaluation sections where the metric is introduced and interpreted.","section":"Section 2.2.2 and Section 3.2"},{"comment":"The downstream age and cognition prediction experiment is a valuable usability check, but it should include a null baseline to interpret the reported metrics. For age, the MAE of about 3.5 years should be compared with the MAE of predicting the mean age; for cognitive scores, a permutation test or a zero-correlation null would clarify whether the reported r values exceed chance. With n=200 and a high-dimensional 1D-CNN input, overfitting is a concern even with cross-validation, and the fold-wise standard deviations are relatively large. Adding these baselines would strengthen the conclusion that predicted connectomes retain subject-specific phenotypic information.","section":"Section 3.4, Table 5"}],"minor_comments":[{"comment":"The text states that training used '8 million streamlines from the training dataset, including 10,000 streamlines per training subject'; with 700 training subjects, 10,000 per subject yields 7 million, not 8 million. Please clarify the discrepancy or correct the number.","section":"Section 2.4"},{"comment":"The abbreviation 'TF AT' in Table 5 appears to be a typo for 'TFAT' (Flanker Attention and Inhibitory Control Test); the abbreviation list in Section 2.1 should be made consistent.","section":"Table 5 and Section 2.1"},{"comment":"The sentence 'The interface was then co-registered to dMRI space using masks provided by the HCP-YA dataset' is unclear; it seems to refer to using HCP-provided transforms or segmentations rather than masks. Please rephrase for clarity.","section":"Section 2.2.1"},{"comment":"The caption states that a Wilcoxon signed-rank test was used to compare intrasubject and intersubject similarity, but these two samples are not naturally paired (each subject has one intrasubject value and many intersubject values). Please specify exactly how the paired test was constructed.","section":"Figure 4 caption"},{"comment":"The reported LERM distances of 7.25±1.35 and 17.03±1.29 should be explicitly identified as intrasubject distances; the text currently introduces them without stating whether they are within-subject or also include a comparison to intersubject values.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competently executed methods study with a clear potential impact, and the authors have been commendably transparent about their evaluation's limitations. The central issue is that the headline correlation needs to be benchmarked against a group-average predictor and supplemented with edge-wise metrics before the subject-specific utility can be accepted. If the authors can provide these analyses and they confirm the current conclusions, the paper would be suitable for publication. The downstream prediction experiment is a good independent check, but it too needs null baselines to be persuasive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a solid engineering paper with a real operational payoff. Instead of running FreeSurfer for hours and then matching streamline endpoints to ROIs, the model takes a whole-brain tractogram (~3M streamlines) and emits two atlas-specific connectomes in about 40 seconds. The multi-task reformulation—classifying streamlines by grey-matter region pair rather than tract name—is the genuinely new piece, and the evaluation is more careful than most in this space: 200 held-out HCP-YA subjects, intrasubject vs intersubject control, test-retest parity, network-measure correlations, and a downstream age/cognition check. The speed and reproducibility claims look credible.\n\nThe soft spots are real but not fatal. The headline r=0.992/0.986 is computed over upper-triangular elements of sparse matrices, so it is dominated by common zero-zero pairs and a few high-count edges. The paper never reports the intersubject correlation values, so we cannot tell how much of that r is subject-specific rather than just 'all healthy young-adult connectomes look alike.' The authors themselves write in the Discussion that 'our evaluation methods may not be sufficiently sensitive, potentially overestimating the true quality of the predictions.' That is a direct self-flagging limitation, and the stress-test concern about a group-average baseline is on point. The downstream age/cognition experiment partially mitigates this—a trivial average connectome would not likely match the traditional connectomes' predictive performance—but the predictive correlations are modest, so this is not a slam-dunk answer.\n\nWhat is missing, more than anything, is a head-to-head comparison with a fast existing pipeline, e.g., FastSurfer followed by tck2connectome. The speed advantage over traditional FreeSurfer is obvious, but the practical gain over current accelerated practice is unquantified. Also, the test-retest cohort's non-overlap with training data is asserted but not explicit, and code/weights are promised but not released.\n\nBottom line: this deserves serious peer review. The central claim—that you can predict two connectomes from tractography in 40 seconds with near-traditional similarity—is supported by the data, with the caveat that the similarity metric needs a baseline and better edge-wise reporting. I'd tell the editor to send it out, and the reviewers to ask for the group-average baseline and a FastSurfer comparison.\n\nRecommendation: accept for peer review.","headline":"DeepMultiConnectome is a genuinely useful engineering contribution for fast multi-atlas connectome generation, but its headline correlation numbers are less decisive than they look—and the authors themselves admit that risk.","tokens_in":15131,"tokens_out":3110,"would_cite":true,"duration_ms":30395,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DeepMultiConnectome predicts structural connectomes directly from diffusion MRI tractography, matching the traditional pipeline at r≈0.99 while cutting the per-subject parcellation bottleneck.","keywords":["structural connectome","diffusion MRI tractography","point cloud neural network","multi-task learning","connectome prediction","parcellation-free connectome","brain network measures","test-retest reproducibility"],"falsifier":"Compute Pearson correlation between predicted and traditional connectomes separately for edges grouped by traditional streamline count, such as the lowest quintile against the highest quintile. If low-count edges show near-zero correlation while overall $r\\approx0.99$ stays high, the headline similarity is carried by high-count edges and per-edge fidelity is not established.","tokens_in":13917,"feed_emoji":"🧠","tokens_out":7476,"duration_ms":70959,"temperature":0.7,"pith_summary":"The paper sets out to prove that a structural connectome can be predicted straight from a whole-brain tractogram, skipping the gray-matter parcellation step that makes the traditional pipeline slow. The proposed DeepMultiConnectome classifies each streamline into the pair of regions it connects, using a shared representation to predict connectomes for an 84-region and a 164-region parcellation at the same time. Across 200 unseen brains, whole connectomes are generated in about 40 seconds and correlate with traditional connectomes at $r=0.992\\pm0.003$ (84 regions) and $r=0.986\\pm0.004$ (164 regions). Test-retest reproducibility and downstream age/cognition predictions are statistically comparable to the traditional method, which is what would make the fast surrogate useful.","feed_headline":"Connectomes from tractography in 40 seconds, no parcellation needed","feed_subtitle":"A multi-task point-cloud model labels 3 million streamlines per brain and rivals the traditional pipeline at r=0.99.","key_machinery":"The load-bearing object is a point-cloud classifier in which every streamline is resampled to 15 evenly spaced 3D coordinates and mapped to a class label standing for an unordered pair of gray-matter regions, plus an unknown class for unassignable streamlines. A multi-task output layer shares the learned streamline representation and predicts the 84-region and 164-region pair labels in parallel; tallying the predicted labels over all streamlines of a tractogram yields the two connectome matrices.","core_discovery":"The central claim is that streamline geometry alone is enough to recover the gray-matter endpoint pair of each fiber, so the entire connectome can be assembled by classifying streamlines rather than by parcellating anatomy. The model emits two connectomes per tractogram in a single forward pass, and the paper reports that these predicted matrices are nearly indistinguishable from traditional ones by Pearson correlation, preserve standard network measures, reproduce across scan sessions as well as the traditional pipeline, and carry comparable subject-specific signal for age and cognition prediction.","pith_inferences":["The shared representation should extend to additional atlases by adding output heads, since all schemes describe the same streamlines; a direct test is whether a third parcellation can be learned without retraining the backbone.","The wide gap between accuracy (~81% and ~70%) and macro F1 (~50% and ~34%) implies that rare, low-count edges are often wrong even though the matrix correlation is high; conclusions about long-range or sparse connections should not be drawn from these matrices until per-edge fidelity is shown.","The evaluation ceiling is the conventional nearest-endpoint labeling, so any systematic bias in that definition of connectivity is inherited exactly; the method cannot be more anatomically correct than its training target.","A practical extension would be to use the classifier's confidence as a per-streamline quality filter within the same 40-second pass, potentially improving downstream connectomes without a separate filtering step."],"forward_implications":["Whole-brain connectomes can be produced for large cohorts without running any anatomical parcellation at inference time.","A single trained model serves multiple parcellation schemes simultaneously, so atlas choice no longer multiplies compute.","Predicted connectomes preserve enough network-level information to support studies of integration, segregation, and resilience comparable to traditional matrices.","Subject-level reproducibility is maintained, so longitudinal or test-retest designs can use the fast surrogate without obvious loss."],"supporting_citations":[{"why":"Supplies the point-cloud streamline representation and model architecture adapted here.","marker":"[52]"},{"why":"Demonstrates efficient point-cloud tractography classification that this method extends to region-pair labels.","marker":"[53]"},{"why":"Provides the tractography and endpoint-labeling tools that generate training labels and the traditional connectomes used as evaluation targets.","marker":"[46]"},{"why":"Defines the cortical parcellations behind the 84-region and 164-region label schemes.","marker":"[17]"},{"why":"Supplies the 1,000-subject dataset plus the 45-subject test-retest scans used for training and reproducibility.","marker":"[47]"},{"why":"Supplies the log-Euclidean Riemannian metric used to compare predicted and traditional connectomes.","marker":"[1]"},{"why":"Recommends the connectome similarity metrics that the evaluation follows.","marker":"[54]"},{"why":"Establishes the Pearson-correlation comparison of connectome matrices used for the headline similarity.","marker":"[34]"}],"fun_headline_variants":["40-second connectomes from raw tractography, no parcellation","Connectome prediction from tractography without gray matter atlases","Single pass returns two connectomes at r=0.99 from streamlines","Predict connectomes directly from dMRI tractography in 40 seconds","Deep learning builds connectomes from streamlines, skips parcellation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That 15 evenly spaced sample points along a streamline carry enough information to identify which pair of gray-matter regions it connects, so a learned geometry-to-label map can replace anatomical parcellation.","fun_headline_variants_meta":{"raw":{"variants":["40-second connectomes from raw tractography, no parcellation","Connectome prediction from tractography without gray matter atlases","Single pass returns two connectomes at r=0.99 from streamlines","Predict connectomes directly from dMRI tractography in 40 seconds","Deep learning builds connectomes from streamlines, skips parcellation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2838,"prompt_tokens":959,"completion_tokens":1879,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1786}},"tokens_in":575,"tokens_out":1879,"duration_ms":12747,"temperature":1.0,"reasoning_tokens":1786,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:44:33.682336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute Pearson correlation between predicted and traditional connectomes separately for edges grouped by traditional streamline count, such as the lowest quintile against the highest quintile. If low-count edges show near-zero correlation while overall $r\\approx0.99$ stays high, the headline similarity is carried by high-count edges and per-edge fidelity is not established.","supporting_citations":[{"cited_title":"J., Makris, N., Rathi, Y., Cai, W., Zhang, F., and O’Donnell, L","cited_arxiv_id":null,"evidence_quote":"Supplies the point-cloud streamline representation and model architecture adapted here."},{"cited_title":"J., Makris, N., Rathi, Y., Cai, W., and O’Donnell, L","cited_arxiv_id":null,"evidence_quote":"Demonstrates efficient point-cloud tractography classification that this method extends to region-pair labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the tractography and endpoint-labeling tools that generate training labels and the traditional connectomes used as evaluation targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the cortical parcellations behind the 84-region and 164-region label schemes."},{"cited_title":"C., Smith, S","cited_arxiv_id":null,"evidence_quote":"Supplies the 1,000-subject dataset plus the 45-subject test-retest scans used for training and reproducibility."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the log-Euclidean Riemannian metric used to compare predicted and traditional connectomes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recommends the connectome similarity metrics that the evaluation follows."},{"cited_title":"A., Parker, D., and Verma, R","cited_arxiv_id":null,"evidence_quote":"Establishes the Pearson-correlation comparison of connectome matrices used for the headline similarity."}],"review_version":1}