{"id":"eb9dbac8-de18-4529-a6ff-9a3494115777","arxiv_id":"2506.11163","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"VeTTA encodes a vascular tree into one vector and recursively decodes it into a geometrically accurate, topologically valid tree, outperforming voxel-based autoencoders on reconstruction metrics.","lead":"The paper introduces VeTTA, a two-stage transformer autoencoder that compresses 3D vessel trees into a single vector and then reconstructs the tree from that vector, including its branching structure and vessel shapes. It reports better reconstructions than voxel-based convolutional autoencoders on a private coronary artery dataset, while needing less GPU memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported superiority over voxel baselines is confounded by input representation: VeTTA receives ground-truth centerlines/radii while baselines receive voxel grids, so Table 2 cannot support the 'superior reconstruction fidelity' claim.","rationale":"The reader's weakest_assumption identifies the same confound I consider most load-bearing. The paper's central claim is two-part: (a) a single vector can encode a full vessel tree and decode to accurate geometry/topology, and (b) this is superior to convolutional approaches. Part (a) is demonstrated by test-set reconstruction, but part (b) is the headline 'superior reconstruction fidelity' and is based on an unfair comparison. A controlled test would settle whether the architecture or the input representation drives the numbers. I considered alternatives: the missing topology metric is a real gap but an omission, not a confound; the proprietary dataset and missing code/error bars weaken reproducibility but do not invalidate the in-paper comparison. The topological guarantee is a genuine architectural contribution, and the memory argument is sound. Thus the appropriate verdict remains CONDITIONAL: the method is plausible and novel, but the central comparative claim needs a fair baseline or explicit qualification. No adjustment to the reader's verdict is needed.","tokens_in":17525,"tokens_out":5856,"duration_ms":64592,"concrete_test":"Fix the input modality: rasterize the exact centerline+radius inputs given to VeTTA into 128^3 voxel grids (e.g., splat cylinders of the stated radii along each centerline), train Conv-3D-AE and Conv-3D-VAE on these grids under the same normalization, latent size (8192), and training schedule, and recompute Table 2 (Dice, Mesh HD, ASD, ACD, APCD). If Conv-3D-AE reaches VeTTA-AE's ~85.4 Dice, the reported superiority is an input-representation artifact. Report bootstrap confidence intervals to check overlap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'superior reconstruction fidelity' rests on Table 2, but the comparison is not controlled. VeTTA is given the ground-truth tree: centerline coordinates, radii, and connectivity (Section 3; Section 4: 'our method first predicts continuous centerline curves (with radius) structured in a tree'). The convolutional baselines receive only 128^3 voxel grids, from which centerlines must be extracted by skeletonization. This is an easier problem setting for VeTTA: the input is already the target representation (a tree), and the main remaining task is compression and reconstruction. The reported advantages (e.g., Dice 85.4 vs 71.8, ACD 0.326 vs 0.587) could therefore be due to the richer input, not to the transformer autoencoder. The paper even states in Section 5: 'Our approach directly leverages explicitly provided curvilinear geometry,' and only acknowledges unfairness in the opposite direction for VesselVAE (Appendix D.4). This confound affects the central quantitative claim, and the absence of error bars or ablations prevents attributing the gap. The topological guarantee of tree-structured output is real but does not address this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VeTTA, a two-stage transformer autoencoder that encodes a vascular tree into a single vector. A vessel autoencoder first maps each vessel segment into an embedding; a tree autoencoder then encodes the whole tree connectivity and geometry into a latent code. A recursive slot-based decoder reconstructs the tree one node at a time, guaranteeing a tree-structured output by construction. Experiments on the public 2D SSA synthetic dataset and a private 3D coronary dataset compare VeTTA with convolutional autoencoders on reconstruction metrics (Dice, Hausdorff, average surface/centerline/point-cloud distances) and show latent interpolations.","tokens_in":17808,"tokens_out":6710,"duration_ms":72290,"significance":"If substantiated, the compact and topology-consistent tree representation could be useful for downstream image-to-geometry tasks and generative modeling of vasculature. The recursive decoder with fixed slots is a sensible design, and the use of continuous centerline geometry is a distinctive contribution. However, the evaluation as presented does not support the headline claim of superior reconstruction fidelity: the comparison against voxel baselines is confounded by input representation, no uncertainty quantification is provided, topology correctness is not measured, and key design choices are not ablated. These gaps are addressable within the scope of a revision.","major_comments":[{"comment":"The claimed 'superior reconstruction fidelity' is not supported by the current comparison because VeTTA and the convolutional baselines receive different input representations. VeTTA is given the ground-truth tree (centerline coordinates, radii, connectivity, and in 3D the pre-trained vessel embeddings), while the baselines receive only 128^3 voxel grids from which centerlines are extracted by skeletonization (Section 4). The improved Dice, ACD, and related metrics in Tables 1 and 2 could therefore stem from the richer input rather than from the transformer autoencoder. Please add a controlled comparison—for example, a variant of VeTTA that takes voxel input, or voxel baselines that are given the same centerline/radius input—or re-frame the conclusion to state that, given an explicit tree representation, VeTTA reconstructs it more accurately than voxel autoencoders reconstruct segmentations.","section":"§4, Tables 1 and 2; §5"},{"comment":"No error bars, confidence intervals, or statistical significance tests are reported for any reconstruction metric. With a single training run per configuration, the differences between VeTTA and the strongly performing Conv-3D-AE baseline (e.g., Mesh HD 2.09 vs 2.22) may not be reliable. Please report the mean and standard deviation across multiple training runs or bootstrap resampling of the test set, and test the main differences.","section":"§5, Tables 1 and 2"},{"comment":"The abstract and Section 5 claim 'accurate topology preservation' and 'topologically correct' reconstructions, but no quantitative topology metric is presented. The recursive decoder guarantees that the output is some tree, but it does not guarantee that the branching pattern matches the ground truth. Please add a quantitative topology evaluation (e.g., branch count error, tree edit distance, or bifurcation match rate) to support the claim.","section":"§5, abstract"},{"comment":"The method depends on many un-ablated design choices: the number of slots (M=32), the matching parameter k=3, the loss reweighting α and position/radius weights, the curvature segmentation sensitivity, the skip-vessel threshold, the evaluation mask m(t), and the Fourier octave choice. The conclusion states that ablation studies are left to future work, which means the current paper does not establish which components drive the reported results. Please include at least a limited set of ablations or sensitivity analyses on the most consequential hyperparameters (slots, k, KL weight, and input feature composition).","section":"§6, Appendix B"},{"comment":"The abstract and introduction state that VeTTA 'substantially lowers GPU memory requirements' compared to 3D convolutional models, but no measured memory usage or computational cost is reported anywhere. Since this is one of the two headline advantages claimed, please provide quantitative memory measurements (e.g., peak GPU memory during training/inference for VeTTA and the baselines) or soften the claim to a qualitative statement about architectural capacity.","section":"§1, §6"}],"minor_comments":[{"comment":"The SSA dataset is referenced only by URL without a formal citation; please provide a bibliographic reference or versioned DOI.","section":"§4"},{"comment":"The caveat that VesselVAE is a generative method and that comparing it on reconstruction metrics is 'somewhat unfair' appears only in Appendix D.4; it should also be stated in the main text where Table 2 is discussed.","section":"§5"},{"comment":"The evaluation mask m(t) in Eq. (1) is introduced without explanation in the main text; please describe its role when it is first used.","section":"§3"},{"comment":"The equation numbering in Appendix A appears to be inconsistent with the main text (e.g., Eqs. (5)-(8) are locally defined); please renumber and unify cross-references.","section":"Appendix A"},{"comment":"The VesselTrees dataset is private and its annotation protocol is not described; please add details on annotation quality, inclusion criteria, and availability, or state that it cannot be released.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The main confound (input representation) is addressable with a controlled experiment, so I do not recommend rejection. However, if the journal expects all comparisons to be apples-to-apples, the current evaluation is not sufficient. I also note that the private VesselTrees dataset makes independent verification of the 3D results difficult; the authors should at least release extracted centerline/radius data or provide a public surrogate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the VeTTA paper.\n\nThe core idea is worth taking seriously. Encapsulating a vessel segment in a first-stage autoencoder and then encoding whole-tree topology with a second transformer, with a recursive decoder that uses fixed slot vectors and clusters them to produce child nodes, is a sensible and novel combination. The topological guarantee is real—because the decoder expands a tree step by step, the output is always a valid tree. That is a genuine strength relative to voxel methods that can produce disconnected artifacts.\n\nThe 2D SSA experiments are on public data, and the method performs on par with or better than convolutional baselines by centerline metrics. The interpolation results are a good qualitative demonstration that the latent space is smoother and more topology-respecting than the baselines'. So the paper demonstrates something useful.\n\nThe problem is the central quantitative claim in Table 2. VeTTA is given ground-truth centerlines, radii, and connectivity as input; the voxel baselines are given 128^3 voxelizations and have to infer centerlines by skeletonization. The reported gains (Dice 85.4 vs 71.8; ACD 0.326 vs 0.587) are therefore not attributable to the architecture alone. The paper notes in Appendix D.4 that comparing VesselVAE on reconstruction is 'somewhat unfair', but does not extend the same logic to its own main comparison. That's an asymmetry worth flagging.\n\nOther issues are minor. There are no error bars or significance tests anywhere. The VesselTrees dataset is private, though the authors promise code release. And there are no ablations; the paper says itself that ablations are future work. Given how many design choices (number of slots, matching parameter k, KL weight, loss reweighting) there are, that's a real gap.\n\nWould I referee this? Yes. The architectural contribution is solid and the evaluation could be fixed: run baselines that also see centerlines, or at least include an ablation that isolates the input representation. The paper should not be accepted as-is, but the method deserves serious consideration.","headline":"Real architectural idea, but Table 2 overclaims: VeTTA's advantage is confounded by giving it centerlines while voxel baselines see only blurry grids.","tokens_in":18359,"tokens_out":2624,"would_cite":true,"duration_ms":28254,"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":"VeTTA encodes a full vascular tree into a single vector and decodes it back into a geometrically accurate, topologically valid tree.","keywords":["tree autoencoders","recursive decoding","coronary artery modeling","vessel trees","vascular networks","transformer autoencoder","topology preservation","latent vector representation"],"falsifier":"Re-run the VesselTrees evaluation with matched inputs: train the convolutional baselines on the same explicit centerline-and-radius data that VeTTA receives, or feed VeTTA only its own 128³ voxelized reconstruction and extract centerlines from it, then compare Dice and centerline-distance scores on the identical test trees. If the Table 2 gaps vanish under either symmetry, the central claim reduces to a statement about input representation rather than about the autoencoder itself.","tokens_in":17261,"feed_emoji":"🌳","tokens_out":13239,"duration_ms":133097,"temperature":0.7,"pith_summary":"The paper claims that an entire vascular tree—its branching pattern together with the continuous geometry of every vessel segment—can be compressed into a single vector and rebuilt faithfully from that vector. VeTTA trains two Transformer autoencoders in sequence: the first learns a compact embedding of individual vessel curves with their radii, and the second pools the tree's edges into one global latent code. A recursive decoder rebuilds the tree node by node, and because every step attaches new child nodes to an already reconstructed node, the output is guaranteed to be a valid tree rather than a disconnected tangle. On a synthetic 2D tree dataset and a 3D coronary artery dataset, the paper reports more accurate reconstructions than voxel-based convolutional autoencoders, substantially lower GPU memory use, and plausible trees from latent-space interpolation. If correct, this gives medical imaging a compact, topologically reliable representation of vessel anatomy for downstream tasks such as image-to-geometry prediction and generative modeling.","feed_headline":"One vector now encodes a full coronary artery tree","feed_subtitle":"The two-stage transformer rebuilds branching vessels from one vector, guaranteeing the output stays a tree.","key_machinery":"The mechanism that carries the argument is recursive decoding conditioned on a global tree code. The latent code $z_t$, produced by pooling transformer-encoded edge features, is concatenated into the edge encodings of the partial tree at each decoding step, so the decoder branch always sees both the full-tree context and the current reconstruction state. A transformer decoder with 32 learned slot vectors outputs per-slot predictions of position, topology, and (in 3D) vessel embedding, log radius, and a skip-vessel flag; agglomerative clustering of these slots in the sinusoidally lifted domain groups several slots into each child node, with the number of children fixed by the query node's predicted topology. Because every appended child connects to an existing queried node, the output cannot contain cycles, which is the paper's topological guarantee. The per-segment geometry itself is carried through the first-stage Vessel Autoencoder and decoded by a small MLP residual model $g_d(z_v, t)$ that refines a linear interpolation between endpoint positions and radii, with an evaluation-time windowing function $m(t)$ that forces exact endpoint matching.","core_discovery":"The central claim is that tree-structured geometry, exemplified by coronary artery networks, admits a single-vector representation from which the original branching structure and continuous vessel geometry can be recovered. VeTTA realizes this in two stages: a Vessel Autoencoder encodes a single vessel segment, represented as 64 samples of position, radius, and arc-length parameter lifted into sinusoidal Fourier features, into a 64-dimensional embedding $z_v$; a Vessel Tree Autoencoder then takes the set of directed edges of the whole tree—each carrying Fourier-lifted endpoint positions, one-hot topology vectors, and (in 3D) the vessel embedding—and pools their transformer encodings into a global code $z_t$. The decoder is the key move: at every recursion step it encodes the partially reconstructed tree with a query flag marking the node to expand, concatenates $z_t$ as context, and uses a transformer decoder over 32 learned slots to predict that node's children, clustering slot predictions in the lifted Fourier domain to form discrete child nodes. A learned start token and a self-loop 'semi-edge' bootstrap the process from an empty partial tree, and the loss is a two-directional matching (right-hand matching plus top-k matching) between slots and ground-truth targets. The authors' claim is that this recursive scheme converts the latent code into a 'program' for rebuilding the tree, giving a structural guarantee of valid tree topology and reconstructions that, in their experiments on a 2D synthetic dataset and a 3D coronary dataset, exceed convolutional autoencoder baselines in Dice score, Hausdorff and centerline distances, and latent interpolation quality.","pith_inferences":["The paper acknowledges an input asymmetry as unfair only for the VesselVAE comparison (Appendix D.4), not for the main voxel-CNN comparison; a matched-input experiment would separate the architecture's contribution from the richer centerline input VeTTA receives.","The authors defer ablation studies to future work, so which components—Fourier lifting, slot clustering, the recursive conditioning, or merely the richer input—drive the reconstruction gains is not yet established.","Because the mechanism assumes only a rooted tree with bounded branching (here at most two children per node), the same two-stage scheme should transfer to other branching anatomies such as the lung airways or hepatic and renal trees, and to any rooted tree-shaped data.","The two-stage split suggests a generative upgrade: place a richer prior over the tree code $z_t$ and sample whole trees whose branch-length and radius statistics could then be validated against the empirical distribution of coronary anatomies."],"forward_implications":["Whole coronary trees become fixed-length vectors, so storing, transmitting, and comparing anatomy no longer requires voxel grids or meshes, and discretization artifacts disappear by construction.","Because the two stages are trained sequentially on large datasets of vessels, the resulting embeddings can be reused as priors for downstream tasks such as predicting vessel geometry directly from cardiac CT images.","Interpolating between latent codes yields intermediate trees that remain valid branching structures, opening a generative route to synthetic-but-plausible coronary anatomies for data augmentation and population shape analysis.","The structural guarantee of tree validity means downstream analyses, such as hemodynamic assessment, start from a connected branching geometry rather than a skeletonized voxel segmentation that may contain spurious branches or gaps.","GPU memory requirements drop substantially relative to 3D convolutional autoencoders, which the authors report makes large-batch training on whole trees feasible on a single 32 GB GPU."],"supporting_citations":[{"why":"Supplies the Transformer architecture used for the vessel encoder, the tree encoder, and the slot-based recursive decoder.","marker":"(Vaswani, 2017)"},{"why":"Provides the sinusoidal Fourier-feature lifting applied to all coordinates, which the paper credits with improving reconstruction accuracy.","marker":"(Tancik et al., 2020)"},{"why":"Defines the GDVM voxel autoencoder that is adapted into the main 3D convolutional baseline.","marker":"(Brock et al., 2016)"},{"why":"Introduces VesselVAE, the closest prior vessel-tree generative model, used as the method-specific baseline.","marker":"(Feldman et al., 2023)"},{"why":"Poisson surface reconstruction converts VeTTA's predicted centerlines and radii into watertight meshes for quantitative evaluation.","marker":"(Kazhdan et al., 2006)"},{"why":"Slot attention underlies the learned-slot set prediction that binds 32 slots to child nodes at each recursive step.","marker":"(Locatello et al., 2020)"}],"fun_headline_variants":["Single vector encodes and rebuilds a vessel tree","VeTTA packs a coronary tree into one vector","One vector reconstructs branching vessel topology","Recursive decoder turns one code into a vessel tree"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison is fair: VeTTA is given the ground-truth centerlines and radii of each vessel as input, while the convolutional baselines receive only voxel grids, so the reported reconstruction advantage could shrink or disappear if every method had to start from the same input representation.","fun_headline_variants_meta":{"raw":{"variants":["Single vector encodes and rebuilds a vessel tree","VeTTA packs a coronary tree into one vector","One vector reconstructs branching vessel topology","Recursive decoder turns one code into a vessel tree"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1370,"prompt_tokens":1028,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":644,"tokens_out":342,"duration_ms":4467,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:33:02.040081+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the VesselTrees evaluation with matched inputs: train the convolutional baselines on the same explicit centerline-and-radius data that VeTTA receives, or feed VeTTA only its own 128³ voxelized reconstruction and extract centerlines from it, then compare Dice and centerline-distance scores on the identical test trees. If the Table 2 gaps vanish under either symmetry, the central claim reduces to a statement about input representation rather than about the autoencoder itself.","supporting_citations":[{"cited_title":"VesselVAE: Recursive Variational Autoencoders for 3D Blood Vessel Synthesis","cited_arxiv_id":"2307.03592","evidence_quote":"Introduces VesselVAE, the closest prior vessel-tree generative model, used as the method-specific baseline."}],"review_version":1}