{"id":"ea2b4d10-f769-4145-8306-99956eb954b7","arxiv_id":"2511.13183","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A generative model that creates entire brain streamline tracts conditioned on whole-brain diffusion MRI, achieving higher precision than prior tractography methods, particularly on degraded data.","lead":"GenTract learns to generate complete white-matter brain pathways directly from diffusion MRI data, using a generative model instead of step-by-step tracking. It reports much higher precision than existing tractography methods, especially on noisy, low-resolution scans that resemble clinical data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training/evaluation circularity: GenTract is trained on PyAFQ's 24-bundle filtered distribution and scored with BundleSeg's atlas; no ablation separates dMRI conditioning from bundle-prior memorization.","rationale":"The reader's weakest assumption identifies the same root cause: training on PyAFQ-filtered, 24-bundle data makes the precision metric overstate biological accuracy. My stress test sharpens this into a concrete missing control: no ablation separates the contribution of dMRI conditioning from the prior learned from the restricted training distribution. If that ablation shows the model ignoring z, the headline result would be an artifact of training/evaluation circularity. The reader's verdict is already CONDITIONAL, and the proposed test would be a reasonable condition for acceptance; therefore I do not move the verdict, but I add a specific experiment that should be run before the claim is accepted.","tokens_in":12477,"tokens_out":4746,"duration_ms":56203,"concrete_test":"Retrain or evaluate the optimal 8-layer Diffusion GenTract on the HCP test set with the conditioning tensor z replaced by (a) all zeros and (b) the population-average z from training subjects, while keeping all other inference settings fixed. If BS % P remains within 10 percentage points of the conditioned model's 61.95%, or the generated tractograms are nearly identical across different test subjects, then the central claim that global dMRI conditioning drives precision is unsupported. Also report BS Bundles in this ablation: if it stays near 36 with BS % P above 50%, the metric is dominated by training-set memorization rather than informative dMRI conditioning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central precision claim is at risk of circularity. GenTract is trained exclusively on PyAFQ-filtered streamlines confined to 24 atlas bundles (Section 5.1), and its headline precision is measured with BundleSeg, an atlas-based tool that labels streamlines by how closely they match a 51-bundle template. A model that has learned to generate only the most canonical training bundles should score high BS % P almost by construction, because the metric rewards exactly the kind of bundle-like streamlines the model was trained to produce. The paper's own Discussion (Section 6) admits that the PyAFQ target leads to a 'constrained bundle distribution' and that proxy evaluation tools 'introduce their own biases,' but the abstract still claims 'complete, anatomically plausible streamlines' and 'whole-brain' tractograms. The low bundle count (36.6/51 vs TractOracle's 48.2 in Table 2) is the flip side of the same coin: the model minimizes false positives by restricting output to the training bundles. The independent TO-Net precision does not fully resolve this, since it also scores geometric bundle plausibility rather than biological ground truth. What is missing is an ablation that tests whether the dMRI conditioning actually drives performance. If a zeroed or fixed conditioning tensor still yields high BS % P, then the advantage is due to memorization of the training bundle prior, not to global generative conditioning.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents GenTract, a conditional generative model for whole-brain tractography. Diffusion-weighted MRI is represented as per-coefficient SH volumes, compressed by per-coefficient VAEs and a class-conditioned encoder into a latent tensor z; a transformer then generates streamline coordinates in parallel, trained with either a diffusion or flow-matching objective. The authors evaluate on HCP test subjects and an external TractoInferno dataset under original, noisy, and low-resolution+noisy conditions, comparing against classical local (iFOD2, SD Stream), RL-based (TractOracle), deep generative local (DDTracking), and classical global (tckglobal) baselines. They report that GenTract achieves substantially higher precision (BundleSeg and TO-Net) at the cost of lower bundle recall, and shows less degradation in low-quality settings.","tokens_in":12952,"tokens_out":5707,"duration_ms":58275,"significance":"The paper introduces a genuinely new paradigm—treating tractography as a single conditional generation step rather than local sequential tracking—and the empirical evaluation is more careful than typical in this area: the main comparison uses two independent evaluation tools (BundleSeg and TO-Net) that were not used to construct training targets, and it includes an external multi-site dataset. If the conditional-generation mechanism is shown to be genuinely driven by the dMRI conditioning, the work would be an important advance in robustness for tractography. The reported precision margins (2.1x over TractOracle on original HCP, order-of-magnitude under LR+noise) are striking. However, the missing conditioning ablation and the constrained training distribution are significant gaps that must be addressed before the central claims can be accepted.","major_comments":[{"comment":"The manuscript does not demonstrate that the anatomical conditioning tensor z actually drives the generated streamlines. All experiments condition on z; no ablation with zeroed, shuffled, or subject-mismatched z is reported. Because the training set itself is a strongly constrained distribution (PyAFQ-filtered 24-bundle tractograms), a model that ignores z and merely samples from the training prior could still score high BS % P in Table 2, since BundleSeg rewards bundle-like, atlas-aligned streamlines. This is load-bearing for the central claim of a 'conditional generative model' and 'direct mapping from dMRI to streamlines.' Please add an unconditional or mis-conditioned baseline (e.g., z=0, z from a different subject) and report the same metrics; without it, the reader cannot distinguish true conditioning from prior memorization.","section":"§4.2 and §5.3"},{"comment":"The abstract and Introduction describe the output as 'complete, anatomically plausible streamlines' and 'whole-brain tractograms,' but the training targets are PyAFQ-filtered tractograms restricted to 24 atlas bundles, and Table 2 shows GenTract recovers only 36.6/51 bundles on the BundleSeg atlas versus 48.2 for TractOracle. The Discussion (Section 6) correctly acknowledges a 'constrained bundle distribution,' but this contradicts the whole-brain framing. Please revise the claims to reflect the limited bundle coverage, or retrain/evaluate on an unfiltered or more complete training target.","section":"§5.1, Table 2, §6"},{"comment":"The conditioning encoder is a central component, but its output z is never validated. No reconstruction quality, latent-space analysis, or evidence that z encodes subject-specific anatomical information is provided. Without such validation, the reader cannot assess whether the global conditioning is meaningful or whether the VAE latents are effectively noise. At minimum, please report reconstruction metrics and show that z varies across subjects and correlates with known anatomy (e.g., bundle presence).","section":"§4.1"}],"minor_comments":[{"comment":"The phrase '1.8x and 2.1x higher than the next-best methods, DDTracking and TractOracle, respectively' is inconsistent with Table 2: DDTracking's BS % P is 0.49 against GenTract's 61.95, a factor of ~126. Please correct this numerical error.","section":"Abstract"},{"comment":"The definition of precision as the percentage of streamlines retained by a filtering tool is acknowledged as a proxy, but the term 'True Positive' is somewhat misleading when no ground-truth tract exists; suggest renaming to 'retained' to avoid implying a biological ground truth.","section":"§5.2"},{"comment":"The text states that 'all streamline coordinates are min-max scaled to [-1,1] using statistics computed from the training set,' but it is not clear whether the same affine transformation is applied independently per subject; please clarify.","section":"§5.1"},{"comment":"The paper repeatedly refers to the Supplementary Material for implementation details and statistical tests, but no supplementary material is included in the arXiv submission. Please ensure the supplementary is provided.","section":"General"},{"comment":"DDTracking's bundle count (9.30) is much lower than other methods; since DDTracking is a deep generative local method, it may not be designed for whole-brain tractography. A sentence explaining this low recall would help.","section":"Table 2"},{"comment":"Model selection is performed on AFQ % P, which is the same filter used for training. Using an independent metric for configuration selection would strengthen the claim that the chosen architecture is not overfit to the training filter.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The conditioning ablation is the key technical gap; if the authors cannot show that z matters, the paper should not be accepted. The whole-brain overclaim should also be fixed. I do not see this as a fundamental circularity that warrants rejection—the external evaluation tools are appropriate—but the current evidence is insufficient to establish the central mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a real idea, not a repackaging. GenTract casts tractography as a conditional generative problem—learn a mapping from whole-brain dMRI to a set of whole streamlines, no seeding mask, no stepwise tracking. The authors compare diffusion and flow-matching backbones and report precision roughly 2x higher than TractOracle on HCP, and an order of magnitude higher under synthetic low-res + noise. More importantly, the main evaluation uses BundleSeg and TO-Net, two tools not used to build the training targets, so the headline number isn't just a self-fulfilling prophecy.\n\nWhat's genuinely good: the robustness story is consistent across internal and external data (TractoInferno), and the qualitative results line up. They also own a key limitation: the training set is PyAFQ-filtered to 24 atlas bundles, which produces a constrained distribution and a high false-negative rate (36.6/51 bundles vs TractOracle's 48.2). That honesty is welcome.\n\nThe soft spot is the one you'd expect. Because the model is trained on that PyAFQ-filtered distribution and the headline metric (BS % P) is an atlas-based score that rewards bundle-shaped streamlines, a model that simply memorizes the training bundle prior could score well even with a zeroed conditioning tensor. The paper never runs that ablation. So we can't yet say the dMRI conditioning is doing the work. The independent TO-Net metric softens this—geometric plausibility isn't identical to atlas matching—but it doesn't fully resolve it. The Discussion acknowledges proxy bias, yet the abstract still says 'complete, anatomically plausible streamlines' and 'whole-brain,' which overstates what a 24-bundle training set can deliver.\n\nOther gaps are smaller: no code or data released, robustness only via synthetic degradation (real clinical validation is future work), and model selection on AFQ % P—the same filter used to build the training targets—does carry some circularity risk, though the independent tools mitigate it.\n\nWho's this for? People working on tractography, especially those interested in generative approaches to medical imaging. It's a legitimate contribution that deserves a proper peer review. I'd recommend accepting it for review, but I'd push hard for the conditioning ablation and a commitment to release code/data before publication.","headline":"GenTract offers a genuinely new global-generative twist on tractography with large reported precision gains, but the design isn't yet clean enough to separate conditioning from bundle-prior memorization.","tokens_in":13306,"tokens_out":3525,"would_cite":false,"duration_ms":34806,"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":"A conditional generative model that samples all white-matter streamlines in parallel achieves 2.1× higher precision than the next-best tractography method, and an order of magnitude more under noise.","keywords":["tractography","generative model","diffusion model","flow matching","white matter","dMRI","streamline generation","global tractography"],"falsifier":"Generate a tractogram on real clinical low-field data and have an independent anatomical expert mark false positives; if GenTract's precision falls to the same level as stepwise methods when scored against a 100-bundle atlas or expert consensus, the reported advantage is an artifact of the restricted training/filtering distribution.","tokens_in":12469,"feed_emoji":"🧠","tokens_out":4635,"duration_ms":45780,"temperature":0.7,"pith_summary":"GenTract reframes whole-brain tractography as a generative task: instead of tracing streamlines step-by-step through local fiber orientation estimates, it learns a direct mapping from the entire diffusion-weighted volume to complete, anatomically plausible streamlines, generating all coordinates simultaneously. The paper claims this is the first generative model for global tractography, and reports precision 2.1× higher than the next-best deep learning baseline on high-quality data, with the gap widening to an order of magnitude on low-resolution, noisy scans, at inference times comparable to other learning-based methods. A sympathetic reader would care because it suggests global tractography—traditionally computationally prohibitive and error-prone—can become practical on clinical-grade data.","feed_headline":"Generative model draws whole-brain tractograms in one pass, doubling precision","feed_subtitle":"On noisy low-resolution scans it beats stepwise tracking by an order of magnitude, making global tractography practical for clinical data.","key_machinery":"The central object is a conditional generative model: a Transformer that takes a learned embedding of the entire dMRI volume and a noise vector, and outputs the coordinates of all streamline points simultaneously. The conditioning tensor is built from per-coefficient autoencoders over spherical-harmonic fiber orientation volumes, fused by a shared class-conditioned encoder, and the Transformer uses self-attention over points along a streamline and cross-attention over the global context. This replaces the stepwise error propagation of local tracking and the energy optimization of classical global tractography with a single parallel sampling process, and removes the need for a seeding mask.","core_discovery":"GenTract learns to sample streamlines directly from a conditional distribution over whole-brain fiber orientation data. A Transformer, conditioned on a learned embedding of the spherical-harmonic coefficients of the dMRI signal, generates the 3D coordinates of all points on a streamline in parallel, producing complete tractograms without any seeding mask. On an independent atlas-based precision metric, it reaches 61.95% versus 28.93% for the next-best baseline, and under low-resolution plus noise it retains 15.73% versus 1.12%, while recovering fewer of the 51 canonical bundles (36.6 vs 48.2). The authors attribute the precision gain to global conditioning and parallel generation, which avoi","pith_inferences":["The reported precision gain likely reflects a strong prior over anatomically plausible bundles learned from the filtered 24-bundle training set; training on a broader, less filtered distribution would test whether the gain survives.","The method's resilience to synthetic noise and resolution loss suggests it captures geometric priors rather than local signal patterns, which may transfer to other dMRI acquisition artifacts such as subject motion.","The same global-conditioning design could be adapted to other streamline-like structure inference problems, such as vascular or neuronal reconstruction.","A decisive validation would be a comparison on real low-field clinical scans, since synthetic degradation may not fully reproduce the characteristics of true clinical noise and blurring."],"forward_implications":["Whole-brain tractograms can be generated in one forward pass, making global tractography fast enough for clinical use.","Precision on noisy, low-resolution scans is an order of magnitude above the best existing baseline, suggesting connectivity mapping is feasible on routine clinical dMRI.","Eliminating the seeding mask removes a major source of operator-dependent variability, improving reproducibility across sites and operators.","Because both diffusion and flow-matching objectives work, the architecture can absorb advances in generative modeling without requiring a new tractography-specific design.","The precision-versus-bundle-recall trade-off points to the training distribution as the limiting factor, not the generative formulation itself."],"fun_headline_variants":["One-pass generative model doubles tractography precision","GenTract: first generative method for global tractography","Generative model for whole-brain tractograms, no seeds","Global tractography without stepwise errors, 3.5x on noise","From dMRI to tractogram: one generative model, all fibers"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The training target is a filtered tractogram restricted to 24 atlas bundles, so the model's high precision is measured against the same type of filtering and may exaggerate biological accuracy if that filtering omits genuine white-matter pathways.","fun_headline_variants_meta":{"raw":{"variants":["One-pass generative model doubles tractography precision","GenTract: first generative method for global tractography","Generative model for whole-brain tractograms, no seeds","Global tractography without stepwise errors, 3.5x on noise","From dMRI to tractogram: one generative model, all fibers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1218,"prompt_tokens":766,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":510,"tokens_out":452,"duration_ms":5050,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:51:25.942232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a tractogram on real clinical low-field data and have an independent anatomical expert mark false positives; if GenTract's precision falls to the same level as stepwise methods when scored against a 100-bundle atlas or expert consensus, the reported advantage is an artifact of the restricted training/filtering distribution.","supporting_citations":[],"review_version":1}