{"id":"88dabf4b-0e9d-4c51-951b-84bdbc8b216a","arxiv_id":"2411.08187","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TractoEmbed fuses streamline, cluster, and patch embeddings to improve white matter tract segmentation, reaching 93.04% accuracy with hyperlocal point clouds.","lead":"TractoEmbed combines three learned embeddings, from streamline, cluster, and patch representations, to label white matter tracts in diffusion MRI. The authors report accuracy gains over prior tract segmentation methods, most notably in a hyperlocal mode that uses a handful of nearby fibers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Table 5 headline Hyperlocal result (93.04/91.38) matches the 'cluster+streamline' ablation at nc=220 (93.038/91.381), not the full three-embedding rows (~92.95/91.28); so the claimed multi-level combination benefit is not internally supported.","rationale":"I read the paper as claiming two things: (1) the modular multi-embedding architecture, especially the addition of patch embeddings, improves tract segmentation; and (2) this yields SOTA accuracy across datasets and age groups. For (1) to hold, the full three-embedding model should be at least as good as its subsets, and the reported 'TractoEmbed' results should correspond to that model. Table 6 contradicts this: in Hyperlocal PCD, cluster+streamline at nc=220 attains 93.038/91.381, while cluster+patch+streamline at the same nc attains 92.837/91.091; even at nc=190 the full model is 92.946/91.284, below the subset. Table 5's 93.04/91.38 is a rounding of the subset row, not the full model. The Local PCD row in Table 5 is not reproducible from any shown Table 6 entry. This is not a matter of external consensus or missing error bars; it is an internal inconsistency in the evidence that directly supports the strongest claim. The reader's atlas-label concern is real for the 'across age groups' generalization statement, but it is secondary: even if the atlas labels were correct, the paper would still need to show that the multi-embedding mechanism, not a two-embedding subset, produces the reported gains. I would not reject the paper outright: the idea of modular embeddings is reasonable, and a subset configuration may still be a valid contribution if clearly identified. But the current reporting conflates the full architecture with the best subset, so the central mechanism is unverified. Thus I keep the reader's conditional-accept posture, with an added condition: identify the exact configuration for every Table 5 row, report seed and error bars, and release the code or provide a reproducible configuration. This concern is fundamentally about internal consistency rather than agreement with literature, so it cannot be waived by appealing to existing consensus.","tokens_in":10654,"tokens_out":9855,"duration_ms":92362,"concrete_test":"Obtain the exact trained configuration used for the 'TractoEmbed(ours)' rows in Table 5 and rerun the 20-subject Hyperlocal PCD test with the full three-encoder architecture described in Section 4.4 at nc=220. If the full model reproduces 93.04/91.38, then Table 6's row labels are wrong; if it reproduces approximately 92.84/91.09, then the headline number comes from the 'cluster+streamline' ablation and the multi-level claim must be revised. Also identify which configuration yields the Local PCD value 92.09/90.07, since no Table 6 row shows it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the multi-level embedding combination, streamline, cluster, and patch encoders concatenated at MECL, is what yields state-of-the-art tract segmentation (Abstract; Section 4.4). Table 5 reports 'TractoEmbed(ours) 93.04 Acc / 91.38 F1' for Hyperlocal PCD. The closest entry in the ablation table, Table 6, is 'cluster + streamline' at nc=220 (93.038/91.381), while the full 'cluster + patch + streamline' model is 92.946/91.284 (nc=190) and 92.837/91.091 (nc=220). Thus the headline number appears to be produced by omitting the patch encoder, and including it slightly lowers both accuracy and F1. The Local PCD row in Table 5 (92.09/90.07) does not match any shown Local PCD row in Table 6 (best shown is about 91.49/89.34), so the configuration for that row is unidentified. This matters because the abstract and Section 4.4 present the three-encoder architecture as the proposed method, and the contribution is explicitly the benefit of combining multiple embeddings. If the full three-embedding model is worse than a two-embedding subset, then the mechanism claimed to explain the SOTA gains is not supported by the paper's own ablation. The atlas-label concern raised by the reader is valid for external generalization, but this internal inconsistency is more load-bearing: it questions whether the reported numbers come from the described method at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TractoEmbed, a modular multi-level embedding framework for white matter tract segmentation. Each streamline is represented at three levels: individual streamline points (15,3), local or hyperlocal point clouds of neighboring streamlines sampled by MDF/FSS, and regional patches obtained by farthest-point sampling and kNN. Three encoders are trained: a CNN on a fiber-descriptor representation of the streamline, a PointNet on the cluster point cloud, and a mini-PointNet+dVAE on patches. Their embeddings are concatenated at a Multi-Embedding Concat Layer and fed to an MLP classifier. Experiments use 1 million streamlines pooled from five public dMRI datasets with whitematteranalysis parcellation labels, and compare against DeepWMA, DCNN++, PointNet, DGCNN, and TractCloud. The paper claims state-of-the-art results, especially in sparse hyperlocal point clouds, and emphasizes modularity for future embedding extensions.","tokens_in":10994,"tokens_out":6728,"duration_ms":64614,"significance":"If the reported results are reproducible and correctly attributed, the framework is a useful contribution: it explores hierarchical streamline representations, reduces reliance on dense global context, and the modular design is conceptually easy to extend. The paper provides detailed architecture tables and an ablation over embedding combinations and point-cloud densities, and it reuses the TractCloud data pipeline to obtain comparable baselines. However, the main empirical claim is weakened by an internal inconsistency between the headline result in Table 5 and the ablation study in Table 6, by the lack of per-dataset validation despite the abstract's generalization claim, and by the absence of uncertainty estimates. In particular, the patch-embedding contribution does not show a consistent benefit in the ablation, which directly undermines the paper's central mechanism.","major_comments":[{"comment":"The headline Hyperlocal PCD result for 'TractoEmbed (ours)' in Table 5 is 93.04% accuracy and 91.38% F1. In Table 6, the closest entries are the 'cluster + streamline' rows at nc=220 (93.038%, 91.381%) and nc=240 (93.020%, 91.346%); the full 'cluster + patch + streamline' rows are 92.946%/91.284% (nc=190), 92.837%/91.091% (nc=220), and 92.876%/91.164% (nc=240). Thus the reported SOTA number corresponds to a two-embedding configuration, not to the three-embedding method described in Section 4.4. In addition, the Local PCD row in Table 5 (92.09%/90.07%) matches no Local PCD row in Table 6; the best shown Local PCD entry is cluster+streamline at nc=190 with 91.494%/89.338%. The authors must state exactly which configuration produced each Table 5 row and explain why the full model is not the one reported.","section":"Table 5 vs Table 6"},{"comment":"The ablation does not support the central claim that combining all three embeddings yields better segmentation. Adding the patch embedding to 'cluster + streamline' decreases accuracy and F1 in four of the six reported columns (hyperlocal nc=220, hyperlocal nc=240, local nc=190, local nc=240) and yields only marginal gains in the other two (accuracy +0.03 or less, with F1 gains up to 0.12). This contradicts the text in Section 5 that 'as the neighboring point cloud becomes sparser ... increasing the need for more embeddings.' Please provide a configuration in which the three-embedding model is consistently superior, or revise the contribution to reflect the two-embedding result.","section":"Table 6"},{"comment":"The abstract claims generalization 'across different datasets' and 'various age groups,' but all results in Section 5 are reported on a single pooled test set of 20 subjects from the combined 100-subject corpus. No accuracy or F1 is given separately for dHCP, ABCD, HCP, PPMI, or BTP. Pooling hides domain shift and does not validate age-group or pathology-level generalization. Please report per-dataset and per-age-group results, or soften the claim accordingly.","section":"Sections 3.2 and 5"},{"comment":"The labels are generated by whitematteranalysis parcellation using an atlas derived from 100 registered HCP young-adult tractograms, and the same atlas is applied to all datasets, including neonatal dHCP and pathological PPMI/BTP. If atlas registration is inaccurate for these populations, the reported accuracy and F1 measure agreement with the atlas rather than anatomical ground truth. The paper should explicitly discuss this limitation and, if possible, evaluate on a subset with expert or alternative labels.","section":"Section 3.1"},{"comment":"All reported numbers are single point estimates from one train/validation/test split; there are no error bars, repeated runs, or significance tests. Many comparisons in Table 5 are within about 1% (e.g., Hyperlocal PCD: 93.04% vs 91.12%), and differences between embedding configurations in Table 6 are often smaller than 0.2%. Without uncertainty estimates, the claim to 'outperform' state-of-the-art methods is not statistically supported. Please add repeated-seed experiments or significance tests.","section":"Tables 5-7"}],"minor_comments":[{"comment":"Table 1 has formatting problems: the dHCP row reads '0 20 vol.' and the column headings (e.g., 'N Volumes (mm3)') are garbled. Please reformat the table so that subject counts, b-values, and volumes are clear.","section":"Table 1"},{"comment":"The FSS radius is reported only as '4mm-6mm'; please state the exact value used for all experiments, since this is a free parameter in the data-preparation pipeline.","section":"Section 3.3"},{"comment":"The paper says 'Code will be made available upon request,' which is not sufficient for reproducibility. Please release the code and the exact subject-wise data splits used for training, validation, and testing.","section":"Section 5"},{"comment":"The symbols 'klocal' and 'khyperlocal' are used interchangeably in several places, and Table 5 uses 'k = 20' and 'k = 5' without a precise definition of which parameter they refer to. Please standardize the notation.","section":"Figure 1 and Section 3.3"},{"comment":"The column header 'TractoEmbed (%)' in Table 7 is confusing because the rows show individual encoders, which are not the full TractoEmbed framework. Please relabel the table to make clear these are single-encoder baselines.","section":"Table 7"},{"comment":"The manuscript contains several grammatical errors, such as 'Diving even further' and 'can further be proven vital.' A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Table 5 and Table 6 is the decisive issue: the headline number is traceable to an ablation row that omits the patch encoder. If the authors cannot identify a configuration that reproduces every Table 5 row with the full three-embedding model, the central contribution needs substantial reframing. The paper appears to be an early preprint version and would benefit from a careful revision before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable incremental extension of TractCloud, clearly written and with a sensible modular design. But the headline result is internally inconsistent with the ablation. Table 5 credits 'TractoEmbed(ours)' with 93.04/91.38 on hyperlocal PCD, and Table 6 shows that exact number is the cluster+streamline ablation at nc=220 (93.038/91.381). The full three-embedding model is 92.946/91.284 at nc=190 and 92.837/91.091 at nc=220. So the best reported number comes from a model that omits the patch encoder, despite the abstract claiming the multi-level combination is the source of the gains. The Local PCD row in Table 5 (92.09/90.07) also does not match any Local PCD row in Table 6, leaving that configuration unidentified. This is not cosmetic; it undermines the paper's central mechanism.\n\nWhat the paper does well: the modular framework is described in enough detail to follow, the data preparation is thorough, and reusing TractCloud's pipeline for baseline comparisons is the right call. The ablation study, though flawed, at least attempts to decompose the contributions. The idea of using hyperlocal point clouds for sparse clinical settings is worth pursuing.\n\nMain soft spots: all results are single runs with no error bars or significance tests; the atlas-derived labels are applied to pediatric and pathological data without independent validation, so the 'generalizes across age groups' claim is overstated; code is 'available upon request,' which is not reproducible enough for a methods paper. The internal inconsistency in Tables 5 and 6 is the most serious issue, because the reader cannot tell which configuration actually produced the headline numbers.\n\nFor whom: tractography researchers interested in modular embedding approaches. The paper deserves a serious referee, but the referee will need to require the authors to clarify the reported configuration, honestly report the full model's performance (even if it is slightly worse), provide error bars or statistical testing, and release code and data. As it stands, the central claim is not supported by the paper's own ablation.","headline":"The modular idea is sensible, but the paper's best reported number comes from a two-embedding ablation, not the full three-embedding model, so the central claim is not backed by the paper's own tables.","tokens_in":11558,"tokens_out":2831,"would_cite":false,"duration_ms":28965,"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":"The paper claims that combining three learned embeddings—from individual streamlines, their local point-cloud neighborhoods, and regional patches—improves white matter tract segmentation from diffusion MRI, especially in sparse…","keywords":["Tract Segmentation","PointCloud","3D Computer Vision","Tractography","Diffusion MRI","Multi-level embedding","Modular framework"],"falsifier":"Take a held-out subset of the pediatric and pathological datasets, have expert neuroanatomists manually label the tracts, and compare TractoEmbed's predictions against those manual labels. If the model's accuracy and F1 against manual labels are not higher than the comparison baselines, or if its errors cluster in tracts where the atlas disagrees with experts, the claim of cross-age generalization would be falsified.","tokens_in":10465,"feed_emoji":"🧠","tokens_out":9407,"duration_ms":81665,"temperature":0.7,"pith_summary":"The paper claims that white matter tract segmentation from diffusion MRI streamlines can be improved by combining three complementary learned embeddings: one capturing a single streamline's shape, one capturing its local neighborhood of nearby streamlines as a point cloud, and one capturing regional patches from that point cloud. The authors argue that this multi-level representation, called TractoEmbed, performs better than existing single-representation and local-global methods, particularly when only a few neighboring streamlines are available, a scenario they call hyperlocal. If the claim holds, tract segmentation would become more accurate and computationally cheaper in focused clinical settings, and the framework's modular design would let researchers plug in new embedding types without retraining the whole system. The paper reports the full model reaching 93.04% accuracy and 91.38% F1 in the hyperlocal setting.","feed_headline":"Three-level embedding fusion pushes brain tract segmentation to 93%","feed_subtitle":"Combining streamline, neighbor, and patch embeddings outperforms existing methods on sparse data.","key_machinery":"The load-bearing object is the Multi-Embedding Concat Layer, which concatenates three embeddings: a 256-dimensional streamline embedding produced by a CNN over a Fiber Descriptor representation; a 1024-dimensional cluster embedding from a PointNet over a hyperlocal point cloud of the streamline plus its five nearest neighbors; and a 1024-dimensional patch embedding from a mini-PointNet and discrete variational autoencoder that reconstructs 64 patches of 16 points sampled by farthest-point sampling and k-nearest neighbors. These three embeddings are concatenated into a 2304-dimensional vector that feeds a classifier MLP. During training, the streamline and patch encoders are frozen after pretraining, while the cluster encoder and classifier are trained end-to-end with focal loss to counter class imbalance.","core_discovery":"On the paper's own terms, the central discovery is that a hierarchy of streamline-level, cluster-level, and patch-level representations, each processed by a task-specific encoder, can be fused into one classifier and outperforms single-representation and existing local-global point cloud baselines in white matter tract segmentation. The reported evaluation shows the full three-embedding model reaching 93.04% accuracy and 91.38% F1 in the hyperlocal point cloud setting, whereas the comparison method reaches 91.12% and 88.66%. The paper further reports that this advantage holds across datasets spanning pediatric, adolescent, adult, Parkinson's, and tumor subjects, all labeled through the same atlas-based parcellation pipeline.","pith_inferences":["A consequence the paper leaves implicit is that its evaluation protocol conflates label consistency with anatomical truth: because every dataset is labeled by the same young-adult-derived atlas, the reported cross-age generalization would only be independently confirmed by manual expert labels on pediatric and pathological brains.","A testable extension would be to measure performance as the hyperlocal neighborhood shrinks from five streamlines to one or two, which would directly probe the method's ceiling for sparse clinical acquisitions.","Adding a rotation-invariant embedding (e.g., based on streamline tangent frames) would be a natural experiment to test whether the current gains persist under arbitrary scanner orientation changes."],"forward_implications":["If correct, tract segmentation becomes effective with only five neighboring streamlines, reducing the data and compute needed for region-specific or time-sensitive analyses.","The modular embedding design means new types of encoders can be added to the multi-embedding layer without reworking the existing components, allowing gradual performance gains as better embeddings are introduced.","The reported improvements in F1 for structurally similar minor and projection fibers would support more reliable connectivity studies in those pathways.","The claimed cross-dataset and cross-age generalization would make the method usable in clinical and developmental studies without per-dataset retraining, assuming the shared atlas label space is valid.","The ablation pattern suggests a practical rule: denser point clouds need fewer embeddings, while sparser clouds benefit from all three, informing how to balance acquisition and modeling."],"supporting_citations":[{"why":"The TractCloud local-global streamline point cloud method is the primary baseline that TractoEmbed compares against in Table 5.","marker":"[30]"},{"why":"DeepWMA supplies the Fiber Descriptor representation and the single-streamline CNN architecture that the streamline encoder is based on.","marker":"[33]"},{"why":"PointNet provides the cluster encoder architecture and also serves as a baseline model in the comparative results.","marker":"[19]"},{"why":"QuickBundles defines the MDF distance used to select local neighboring streamlines for cluster data.","marker":"[9]"},{"why":"Fast Streamline Search (FSS) is used to retrieve the five hyperlocal streamlines that form the hyperlocal point cloud.","marker":"[21]"},{"why":"The anatomically curated fiber clustering white matter atlas is the label source that defines ground truth for all training and evaluation.","marker":"[32]"},{"why":"The discrete variational autoencoder (dVAE) provides the reconstruction objective and architecture for the patch encoder.","marker":"[20]"}],"fun_headline_variants":["Three-level embedding fusion lifts tract segmentation to 93%","TractoEmbed's hierarchical embeddings beat prior tract segmentation","Fusing three embeddings improves brain tract segmentation","Modular embedding framework sharpens white matter tract segmentation","Multi-scale embeddings boost tract segmentation accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline treats the atlas-based parcellation as ground truth for all datasets; if the atlas, which was derived from young-adult HCP brains, mislabels developing or pathological brains, then the reported accuracy and F1 measure agreement with the atlas rather than anatomical truth, and the claimed generalization across age groups is not independently validated.","fun_headline_variants_meta":{"raw":{"variants":["Three-level embedding fusion lifts tract segmentation to 93%","TractoEmbed's hierarchical embeddings beat prior tract segmentation","Fusing three embeddings improves brain tract segmentation","Modular embedding framework sharpens white matter tract segmentation","Multi-scale embeddings boost tract segmentation accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00101,"raw_usage":{"total_tokens":4199,"prompt_tokens":807,"completion_tokens":3392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":3319}},"tokens_in":423,"tokens_out":3392,"duration_ms":24174,"temperature":1.0,"reasoning_tokens":3319,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:53:07.847673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out subset of the pediatric and pathological datasets, have expert neuroanatomists manually label the tracts, and compare TractoEmbed's predictions against those manual labels. If the model's accuracy and F1 against manual labels are not higher than the comparison baselines, or if its errors cluster in tracts where the atlas disagrees with experts, the claim of cross-age generalization would be falsified.","supporting_citations":[{"cited_title":"IEEE transactions on medical imaging38(8), 1910–1922 (2019)","cited_arxiv_id":null,"evidence_quote":"The TractCloud local-global streamline point cloud method is the primary baseline that TractoEmbed compares against in Table 5."},{"cited_title":"Neuroimage179, 429–447 (2018)","cited_arxiv_id":null,"evidence_quote":"DeepWMA supplies the Fiber Descriptor representation and the single-streamline CNN architecture that the streamline encoder is based on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PointNet provides the cluster encoder architecture and also serves as a baseline model in the comparative results."},{"cited_title":"Frontiers in neuroscience6, 175 (2012) TractoEmbed 15","cited_arxiv_id":null,"evidence_quote":"QuickBundles defines the MDF distance used to select local neighboring streamlines for cluster data."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"The discrete variational autoencoder (dVAE) provides the reconstruction objective and architecture for the patch encoder."}],"review_version":1}