{"id":"0dcd07e5-4a7e-4cbf-83b6-ea09c2e8029e","arxiv_id":"2505.13318","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An autoregressive GPT-2 model was trained on VQ-VAE tokens to generate new 3D blood vessel trees from the Aneurisk dataset.","lead":"Doctors and engineers need realistic 3D models of blood vessels, and this paper shows that a language model can be trained to invent new vessel shapes from examples. The key idea is to encode a vessel tree as a sequence of symbols, let the model learn how the symbols usually follow each other, and then decode new sequences back into 3D shapes with realistic cross-sections.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3's post-hoc 1e-2 absent-node threshold is an untested topology fix; if it misclassifies nodes, every generated vessel inherits corrupt tree structure.","rationale":"The reader's verdict was CONDITIONAL, with the weakest assumption being exactly the absent-node threshold. This stress-test agrees. The strongest claim is that VesselGPT effectively captures intricate branching patterns. For that to hold, the discrete vocabulary must encode tree topology reliably. The paper openly states that zero vectors map to different token sequences and that a hard threshold is applied after decoding, yet no topology-specific metric is reported. This is an untested, load-bearing assumption: it does not merely affect numeric scores, it determines whether each generated object is a valid vascular tree at all. The concern is supported by the manuscript's own implementation note, so it is not manufactured. At the same time, the paper provides code, a clear two-stage architecture, and self-reported reconstructions, so the appropriate disposition is to keep the verdict CONDITIONAL rather than reject: a targeted threshold-sweep experiment could confirm or refute the robustness. Therefore verdict_should_be is UNCHANGED, and agreement_with_reader is agree.","tokens_in":7986,"tokens_out":5294,"duration_ms":49179,"concrete_test":"Using the released code, reconstruct every held-out Aneurisk tree with the frozen VQ-VAE and sweep the absent-node threshold over {1e-4, 1e-3, 1e-2, 5e-2, 1e-1}; for each threshold compute node-existence F1 and exact tree-topology accuracy against ground truth. Also sample 1,000 sequences from the trained GPT-2 and count how many thresholded decodings form valid binary trees (no spurious children, no missing branches). If F1 at the default 1e-2 is below 0.95 or varies materially across the sweep, the post-hoc threshold does not reliably preserve topology and the generation claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that binary-tree topology survive VQ-VAE encoding/decoding and GPT-2 generation. Section 3 (Implementation details) concedes that zero placeholder vectors for absent child nodes \"are often mapped to different token sequences\" and that the authors therefore \"threshold values lower than 10^-2 to zero\" after decoding to restore non-existing nodes. This threshold is a post-hoc heuristic, not a learned invariant. It assumes a global scalar cutoff cleanly separates absent nodes from real nodes in a normalized [-1,1] attribute space that includes coordinates and B-spline coefficients/control points; a genuine thin branch or small spline weight below 1e-2 could be erased, while a decoded absent node with small nonzero residual could be kept. No experiment measures topology reconstruction accuracy (node-existence precision/recall or exact tree-structure match) or the validity of generated trees after thresholding. Since all meshes are built from the thresholded tree, any topology error propagates directly into the generated geometry, and the reported point-cloud metrics and length/tortuosity cosine similarities do not isolate this failure mode. This is the load-bearing weak point of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VesselGPT, a two-stage generative model for 3D vascular trees. Stage 1 trains a vector-quantized autoencoder (VQ-VAE) to map a preorder-serialized sequence of vessel nodes into a discrete codebook; each node encodes 3D coordinates and B-spline cross-section parameters, with zero vectors marking absent child nodes so that the binary tree structure can be restored. Stage 2 trains a GPT-2-style decoder-only transformer to predict codebook indices autoregressively. A generated token sequence is decoded by the VQ-VAE decoder, and a centerline-fitting / SDF / marching-cubes pipeline turns the decoded parameters into a mesh. Experiments on the Aneurisk dataset use 528 binary trees obtained by augmentation from 100 meshes, with quantitative evaluation via MMD, COV, 1-NNA, and cosine similarity of length and tortuosity histograms. The authors report high-fidelity reconstruction and claim that this is the first autoregressive method for blood vessel generation.","tokens_in":8215,"tokens_out":4561,"duration_ms":40134,"significance":"VesselGPT is a sensible and potentially valuable application of discrete token-sequence modeling to anatomical trees. The paper has clear strengths: it uses a public dataset, releases code, proposes a compact discrete representation, and explicitly models non-circular cross-sections with B-splines, which is a genuinely useful departure from the common single-radius parameterization. If the topology-reconstruction and evaluation issues raised below are resolved, the method would be a meaningful contribution to generative geometry and medical image computing. At present, however, the quantitative evidence is partially inconsistent with the paper's claims, and the robustness of the binary-tree reconstruction is not established.","major_comments":[{"comment":"The fixed 1e-2 threshold used to restore absent child nodes is load-bearing and is not validated. The text concedes that zero placeholder vectors 'are often mapped to different token sequences' and that values below 10^-2 are thresholded to zero after decoding, but no experiment reports node-existence precision/recall, exact tree-structure match, or validity of thresholded generated trees. Since every mesh is built from the thresholded binary tree, any misclassification of thin branches or small spline weights directly corrupts the generated topology. Please add a quantitative topology-reconstruction evaluation, a sensitivity analysis for the threshold, or replace the heuristic with an explicit existence bit/token so that node absence is learned rather than assumed.","section":"Section 3, Implementation details"},{"comment":"Table 1 is incomplete and therefore cannot support the comparative claims: the 1-NNA entries for Sinha et al. appear as truncated '0.' values for both heights, and no standard deviations, confidence intervals, or significance tests are reported for any method. Because the claims about outperforming baselines rest on this table, the authors should either provide the missing entries or remove the corresponding comparative statements, and should report repeated-generation statistics with multiple seeds.","section":"Table 1"},{"comment":"The quantitative results are mixed with respect to the stated advantages. At height 15, VesselVAE achieves lower MMD (0.014 vs 0.14) and higher COV (0.49 vs 0.31); at height 20 it achieves lower MMD (0.013 vs 0.14) and comparable COV (0.45 vs 0.41). The paper explains this only qualitatively, and the 1-NNA interpretation is confusing because the caption says the optimal value is 0.5 while the text appears to treat lower values as better. Please report a properly interpreted 1-NNA comparison and either demonstrate a statistically significant advantage on the primary metrics or reframe the claims accordingly.","section":"Table 1 and Section 4"},{"comment":"The description of point-based metric computation is ambiguous: 'we randomly sample 1000 points from all baseline generated meshes' suggests that the reference set used for MMD/COV/1-NNA is not specified. Additionally, no train/validation/test split or evaluation protocol is described for the 528 trees, so it is unclear whether the reported numbers are on held-out samples. Please specify the exact protocol for both the point-based and vessel-specific metrics.","section":"Section 4, Metrics"}],"minor_comments":[{"comment":"The abstract contains the typo 'previous' methods' instead of 'previous methods'; please correct the apostrophe and grammar.","section":"Section 1, Abstract"},{"comment":"The sentence 'but faces limitations with deeper trees and simplifies radius parametrization to a single value per centerline point, loosing fine details' uses 'loosing' instead of 'losing'; please revise.","section":"Section 4"},{"comment":"The phrase 'could enhance the and realism of the generated structures' is grammatically incomplete; please revise to 'could enhance the realism'.","section":"Section 5, Conclusions"},{"comment":"The codebook size K and the VQ-VAE commitment weight lambda are not reported in Implementation details, even though they are central to reproducing Stage 1; please add these values.","section":"Section 3, Implementation details"},{"comment":"The renders in Figure 2 lack scale bars and clear axis information, which makes it difficult to compare tree sizes across methods; please add consistent scale information.","section":"Figure 2"},{"comment":"The claim 'this work is the first to generate blood vessels in an autoregressive manner' is a strong novelty statement; please clarify the scope relative to prior sequential or token-based generative models, since the comparison with VesselVAE and diffusion baselines does not directly address this claim.","section":"Section 1, Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal and presents a reasonable pipeline, but the load-bearing threshold heuristic and the incomplete, non-significant quantitative comparison need to be fixed before the central claims can be accepted. The comparison with the authors' own VesselVAE is a legitimate baseline, and I do not see a circularity problem. The manuscript would also benefit from a careful proofread to remove the several typos."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a legitimately new application—the first autoregressive generator for vascular trees, combining VQ-VAE tokenization with GPT-2 and B-spline cross-sections. Code is public, the dataset is standard Aneurisk, and the qualitative results look plausible. If you work on medical image synthesis or anatomical tree generation, this is worth paying attention to.\n\nWhat's actually new: the discrete vocabulary for vessel nodes, preorder serialization with zero placeholders for absent children, and the B-spline parameterization of cross-sections that goes beyond the usual centerline-plus-radius. The idea is clean and the paper is mostly readable. They also show the method scales to deeper trees than VesselVAE, which is a fair point. Citations are appropriate; comparing against their own VesselVAE is natural since it's the closest prior work.\n\nThe soft spots are in the evaluation, not the core idea. Table 1 is incomplete—Sinha's 1-NNA entries are missing—and there are no error bars or significance tests. The quantitative comparison is genuinely mixed: VesselVAE has lower MMD and higher COV at both heights, and VesselGPT only wins on some 1-NNA numbers. The authors argue their B-spline representation captures non-circular cross-sections that standard metrics miss, but they never measure that claim. That's a central advantage asserted qualitatively, which is a real gap.\n\nThe bigger technical concern is the absent-node threshold in Section 3. They admit zero placeholder vectors encode to inconsistent tokens, so after decoding they threshold values below 1e-2 to zero to restore absent children. This is a post-hoc fix. In a normalized [-1,1] space with coordinates and spline weights, a global cutoff could erase a genuine thin branch or keep a spurious one. They report no topology reconstruction accuracy, so we don't know how often this happens. Since meshes are built from the thresholded tree, any error propagates into every generated sample. This is the load-bearing weak point and needs a direct experiment.\n\nThat said, the paper is not broken; the threshold may work fine in practice, we just can't tell. The right response is revision, not rejection. Add complete tables with variance, topology precision/recall or exact-match, and an ablation of the cross-section representation. If those hold up, this will be a useful contribution.\n\nWho is this for? Medical imaging researchers who want synthetic vascular trees for augmentation or simulation, and anyone working on autoregressive generation of structured 3D data. It deserves a serious referee. I'd engage with it, but I'd want the missing numbers before trusting the results.","headline":"First autoregressive vascular-tree generator, with a clean VQ-VAE/GPT-2 pipeline and B-spline cross-sections; evaluation gaps and an unvalidated absent-node threshold keep it from being trustworthy yet.","tokens_in":8758,"tokens_out":3211,"would_cite":true,"duration_ms":26788,"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":"Blood-vessel trees can be synthesized autoregressively by tokenizing vessel geometry and sampling new trees token by token with a transformer.","keywords":["autoregressive generation","vascular geometry","blood vessel synthesis","VQ-VAE","tokenization","B-spline cross-sections","anatomical trees","3D mesh generation"],"falsifier":"Take a set of vessel trees with known topology, encode and decode them through the VQ-VAE, and count how often the fixed 1e-2 threshold misclassifies an absent child as present or vice versa; if the error rate is non-negligible, topology errors are being silently injected into every generated tree.","tokens_in":1538,"feed_emoji":"🩸","tokens_out":1799,"duration_ms":83103,"temperature":0.7,"pith_summary":"VesselGPT is designed to show that blood-vessel trees can be generated as ordered token sequences rather than through hand-coded branching rules. The central claim is that a VQ-VAE can convert vessel-node geometry into a learned discrete codebook, after which a GPT-2-style transformer can generate new trees by predicting the next codebook token. The representation differs from earlier centerline-and-radius methods because each node also carries B-spline weights describing the vessel cross-section, so non-circular shapes and aneurysms can be captured. The paper reports that generated trees closely match real vascular distributions in tortuosity and total length, and that the pipeline reconstructs high-fidelity geometry from compact discrete tokens. A sympathetic reader would take this as evidence that autoregressive token prediction is a viable route to anatomical tree generation.","feed_headline":"Blood-vessel trees generated token by token, not by hand-coded rules","feed_subtitle":"VesselGPT turns vessel geometry into a learned codebook and samples new branching trees from it.","key_machinery":"The machinery is a two-stage token pipeline. Stage one is a vector-quantized autoencoder whose codebook replaces raw continuous node attributes, namely coordinates plus B-spline coefficients and control points, with discrete embedding indices; the encoder and decoder each contain a transformer layer, and training uses a reconstruction loss plus commitment losses with a stop-gradient operator. Stage two is a decoder-only GPT-2 transformer trained to maximize the log-probability of the next codebook index conditioned on previous tokens, then sampled with beam search from a start token to a stop token. The decoded token sequence is turned into a mesh by fitting B-splines along each centerline branch, building a signed distance field by interpolating cross-sections, and applying marching cubes.","core_discovery":"On the paper's own terms, the finding is that vascular anatomy can be reduced to an ordered sequence of discrete tokens without losing its branching structure: each binary tree is serialized in preorder, absent children are marked by zero vectors, and a VQ-VAE learns a codebook such that the sequence of codebook indices can be predicted autoregressively. Generated token sequences are decoded back into spline coefficients, and a signed-distance-field meshing step turns them into 3D surfaces. The cross-sectional B-splines are the key to preserving morphology that a single radius value would discard. The paper presents this as the first autoregressive blood-vessel generator, with reconstruction and distributional matches obtained on a public dataset of 100 patient-derived vessel segments.","pith_inferences":["One direction the paper leaves open is conditional generation: conditioning the same GPT-2 backbone on a requested length, caliber, or diagnosis could sample trees for a specific clinical context, but the paper does not explore this.","The fixed threshold for identifying absent children is a silent risk in deeper trees; a robustness study across vessel sizes would show whether small decoding errors accumulate into topology corruption.","If the tokenization transfers, the two-stage design could be trained on other tubular structures such as airways, ducts, or plant roots, since the method itself is not vessel-specific.","The B-spline cross-sections offer a natural path to pathology generation: encoding aneurysm-like bulges as codebook patterns could synthesize pathological variants for data augmentation, though the paper stops at morphology matching and does not validate clinical utility."],"forward_implications":["A compact discrete codebook (16 tokens per node, 64-dimensional vectors) is enough to reconstruct vessel geometry at high fidelity, so vascular models could be stored or transmitted as token streams.","Because cross-sections are represented by B-splines instead of a single radius, shapes that deviate from circular tubes, including aneurysms, are representable in the generated output.","The pipeline scales to tree heights that break the recursive baseline, since the transformer processes the full serialized sequence rather than recursing level by level.","Generated trees match real data on tortuosity (cosine similarity 0.97) and total length (0.88), which the paper reads as evidence of anatomical realism beyond point-cloud metrics.","The same preorder serialization and zero-vector placeholder scheme could apply to other anatomical trees such as airways and renal tubules, which the paper names as targets."],"supporting_citations":[{"why":"Supplies the VQ-VAE objective and codebook-quantization mechanism that turns vessel nodes into discrete tokens.","marker":"[31]"},{"why":"Provides the GPT-2 decoder-only transformer architecture used for autoregressive next-token prediction.","marker":"[22]"},{"why":"Is the public dataset of 100 patient-derived vessel segments from which the training trees are built.","marker":"[2]"},{"why":"Is the recursive variational autoencoder baseline whose centerline-and-radius representation and depth limits this work aims to surpass.","marker":"[9]"},{"why":"Is the diffusion baseline compared on MMD, COV, and 1-NNA for generated tree quality.","marker":"[27]"},{"why":"Is the marching cubes algorithm used to extract the final vessel surface mesh from the signed distance field.","marker":"[18]"},{"why":"Is the transformer layer embedded in the VQ-VAE encoder and decoder to provide context-aware node encoding.","marker":"[32]"},{"why":"Is the stop-gradient technique used to train the discrete codebook in the VQ-VAE loss.","marker":"[3]"},{"why":"Defines the MMD, COV, and 1-NNA point-cloud metrics used to benchmark generation quality.","marker":"[35]"}],"fun_headline_variants":["VesselGPT: autoregressive vascular trees from learned codebooks","GPT-2 predicts vessel tokens to synthesize branching geometries","Vascular trees as language: discrete tokens, autoregressive growth","First token-based autoregressive model for blood vessels","Codebook tokens serialize vessel splines for GPT-2 synthesis"],"cache_read_input_tokens":10880,"weakest_assumption_plain":"The load-bearing premise is that a fixed numerical cutoff of 1e-2 cleanly separates absent tree nodes from real ones after VQ-VAE decoding, even though the authors note that zero placeholder vectors are often mapped to inconsistent token sequences.","fun_headline_variants_meta":{"raw":{"variants":["VesselGPT: autoregressive vascular trees from learned codebooks","GPT-2 predicts vessel tokens to synthesize branching geometries","Vascular trees as language: discrete tokens, autoregressive growth","First token-based autoregressive model for blood vessels","Codebook tokens serialize vessel splines for GPT-2 synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3389,"prompt_tokens":851,"completion_tokens":2538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2457}},"tokens_in":467,"tokens_out":2538,"duration_ms":16695,"temperature":1.0,"reasoning_tokens":2457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:15:06.252420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of vessel trees with known topology, encode and decode them through the VQ-VAE, and count how often the fixed 1e-2 threshold misclassifies an absent child as present or vice versa; if the error rate is non-negligible, topology errors are being silently injected into every generated tree.","supporting_citations":[{"cited_title":"Web Site (2012)","cited_arxiv_id":null,"evidence_quote":"Is the public dataset of 100 patient-derived vessel segments from which the training trees are built."},{"cited_title":"In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Is the recursive variational autoencoder baseline whose centerline-and-radius representation and depth limits this work aims to surpass."},{"cited_title":"$TrIND$: Representing Anatomical Trees by Denoising Diffusion of Implicit Neural Fields","cited_arxiv_id":"2403.08974","evidence_quote":"Is the diffusion baseline compared on MMD, COV, and 1-NNA for generated tree quality."},{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Defines the MMD, COV, and 1-NNA point-cloud metrics used to benchmark generation quality."}],"review_version":1}