{"id":"8f5f3508-159b-4398-ace7-aa0dee11dfc2","arxiv_id":"2509.03095","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TRELLIS-derived surface features improve aneurysm classification, segmentation, and hemodynamic simulation, including a 15% lower blood-flow prediction error.","lead":"TRELLIS, a 3D generative model trained on everyday objects, is used as a feature extractor for brain aneurysm data. Adding its features improves aneurysm classification, segmentation, and blood flow simulation, cutting simulation error by about 15%, which could help in data-scarce medical imaging.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: reported gains may stem from 1024D feature dimensionality or global object cues rather than TRELLIS's pretrained cross-domain transfer; no non-transfer high-dimensional baseline is tested.","rationale":"The reader's weakest assumption is exactly the confound I identify: the paper does not test whether improvements come from transferable geometric knowledge or from the much higher dimensionality of the TRELLIS features and global object cues. I agree this is the most load-bearing issue because the paper's conclusion is framed around cross-domain transfer from a generative model, not just empirical gains. The reader also noted missing classification split details and the gap between the paper's own baselines and published baselines; those are important protocol concerns, but the dimensionality/transfer confound is the one that, if unresolved, undermines the mechanistic claim. I would keep the verdict CONDITIONAL, requiring the control ablation before accepting the transfer interpretation. Since the reader already arrived at CONDITIONAL, no change in verdict is needed.","tokens_in":13724,"tokens_out":5553,"duration_ms":71005,"concrete_test":"Rerun the classification and segmentation experiments of Sections 6.1–6.2 replacing the TRELLIS features with a fixed, randomly initialized 1024-dimensional descriptor computed from the same local inputs: for each point, apply a seeded random Fourier feature map to (x, y, z, nx, ny, nz), then train the same PointNet, PointNet++, and MLP-only models with identical splits, hyperparameters, and epochs. If this non-transfer control reaches or substantially narrows the gap to the TRELLIS-augmented models in Tables 1–2, the gains cannot be attributed to pretrained cross-domain knowledge. If it does not, the transfer interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that latent geometric knowledge pretrained on non-medical 3D data transfers to aneurysm analysis. The supporting experiments in Sections 4.2/4.3 and 6.1/6.2 compare models using 3D normals against models using 1024D TRELLIS features, confounding feature dimensionality with pretraining/transfer. TRELLIS features are produced by a global encoder that sees the entire object through rendered views, so they may encode object-level cues such as bounding-box shape, orientation, or vessel context, rather than reusable local surface geometry. The MLP-only ablation in Section 4.2, which excludes spatial coordinates, does not resolve this: a high-dimensional input can memorize object-level statistics even without pretrained semantic structure. No control uses an equally high-dimensional non-transfer descriptor (e.g., random Fourier features of coordinates/normals, or an untrained TRELLIS encoder). Without such a control, the 'cross-domain feature-transfer' interpretation in the abstract and conclusion is not established; the observed improvements could be due to increased capacity or global cues rather than transferable geometric knowledge. This is load-bearing because the paper's contribution is specifically the transfer mechanism, not merely better performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using latent features from the pretrained TRELLIS 3D generative model as additional point/mesh descriptors for three downstream tasks: classification of aneurysms versus healthy vessels on Intra3D, segmentation of aneurysm/vessel regions on the same dataset, and prediction of time-evolving hemodynamics on the AnXplore dataset. The authors replace point normals (or augment mesh features) with 1024-dimensional TRELLIS features, and report large improvements in classification accuracy/F1, segmentation IoU/DSC, and a roughly 15% reduction in simulation RMSE. The paper also includes feature-space visualizations, a clustering analysis, and an MLP-only ablation, and makes the code publicly available.","tokens_in":14026,"tokens_out":7909,"duration_ms":86421,"significance":"The paper targets a timely and practically relevant question: whether features from a large-scale 3D generative model can compensate for scarce medical data. It evaluates the idea on three tasks and reports consistent improvements; the public code and multi-seed reporting are strengths. I agree there is no circularity: the TRELLIS encoder is fixed and was not trained on the target labels. However, the specific contribution claimed—cross-domain transfer of geometric knowledge—is not yet demonstrated because input dimensionality and global object cues are confounded with pretraining. If the requested controls confirm the transfer effect, the paper would be a useful addition to the medical 3D deep-learning literature.","major_comments":[{"comment":"The central claim is that pretrained TRELLIS geometry transfers to medical 3D tasks, but the experiments confound transfer with input dimensionality and global context. The baseline uses coordinates + normals (3D per point), while the TRELLIS condition uses 1024-dimensional per-point features. The MLP-only ablation (Section 4.2) excludes coordinates but still feeds 1024D features, so it does not rule out memorization of object-level statistics. Moreover, the TRELLIS encoder is a global, multi-view model; its per-voxel tokens may encode orientation, bounding-box shape, or global vessel context rather than reusable local surface geometry. Please add at least one non-transfer high-dimensional control, e.g., random Fourier features of coordinates/normals projected to 1024D, a random/untrained TRELLIS encoder, or PCA/random projections of normals to 1024D. The same control is needed for the G","section":"Section 4.2, Table 1"},{"comment":"The classification protocol is under-specified. The text says 'training 5 times the models over 5 different seeds' but does not describe the train/validation/test split. If evaluation is on the training set, the 99.88/100 accuracy values are uninterpretable. Please state the split, whether it is stratified by class, and whether all models share identical point sampling. In addition, the authors' PointNet++ baseline without TRELLIS features reaches F1 ≈ 0.81 and aneurysm accuracy ≈ 52%, whereas the published PointNet++ results from Yang et al. (2020) on the same benchmark report F1 ≈ 0.89 and aneurysm accuracy ≈ 86.7% at 512 points. This large gap suggests protocol or implementation differences; the claimed 'state-of-the-art' comparison must be made under an identical protocol, including the same training setup and preprocessing.","section":"Section 6.1, Table 1"},{"comment":"The description of the classification dataset is ambiguous: '1,694 healthy vessels and 215 aneurysms, which can be combined with the 116 annotated aneurysms used to train segmentation models, resulting in a total of 331 aneurysms for classification.' In the original Intra3D dataset, the 116 annotated segmentation cases are a subset of the 215 classification cases. If that is the case, combining them double-counts 116 cases; if they are disjoint, that must be stated explicitly. This affects the composition of every classification result in Table 1.","section":"Section 2.1"}],"minor_comments":[{"comment":"Caption says 'results are ... F1-score' but the table reports IoU and DSC; correct the caption or the table.","section":"Table 2 caption"},{"comment":"Caption says 'vessels segment accuracy' and 'aneurysms segment accuracy' for a classification task; rename to 'vessel accuracy' and 'aneurysm accuracy'.","section":"Table 1 caption"},{"comment":"The statement 'the optimal number of clusters for the 101 aneurysms was 15' gives no method (silhouette, elbow, etc.) for choosing k; please specify.","section":"Section 3.2"},{"comment":"The discussion of reducing rendering views from 200 to 100/50 is speculative and not backed by experiments; either add an ablation or clearly label it as future work.","section":"Section 6.4"},{"comment":"The citation 'A et al., 2024' should be expanded to a proper author list (Goetz, A. et al.) in the text, and the reference formatting should be consistent.","section":"References"},{"comment":"The abstract says 'replacing conventional point normals or mesh descriptors', but Section 5.2 states that TRELLIS features are 'combined with the original mesh features' for the GNN experiments. Please make the description consistent.","section":"Abstract / Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"No concerns about circularity or citation behavior: the TRELLIS encoder is fixed and not fitted on target labels. The main issue is that the transfer claim is confounded and the classification protocol is incomplete. The paper is likely to be salvageable with the requested controls and protocol details, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper to know: they take TRELLIS, a large 3D generative model trained on Objaverse and similar, and use its per-voxel 1024-d features as surface descriptors for aneurysm classification, segmentation, and blood-flow GNN prediction. On Intra3D classification, a plain MLP on TRELLIS features gets ~99% F1, beating published PointNet++ and 3DMedPT. That's the real news: a frozen generative encoder, never trained on medical data, yields features that almost trivialize the classification task. That empirically holds.\n\nWhat's good: the idea is clearly stated, the pipeline is simple, and the code is public. They report means and stds over seeds and folds, and they honestly discuss the 5-minute-per-object encoding cost. The feature-space analysis (PCA/t-SNE/clustering) is appropriate and shows the features cleanly separate aneurysm from vessel. The GNN simulation error reduction (15%) is consistent with the rest.\n\nSoft spots, in order:\n\n1. The classification protocol is under-specified. They say the 215 classification aneurysms \"can be combined with the 116 annotated aneurysms... resulting in a total of 331.\" If the 116 are a subset of the 215, that's double-counting; if they are separate, the numbers need clarification. The paper never says which split was used (random? stratified? whether the 116 overlap the 215). That's load-bearing for trusting the numbers.\n\n2. The baseline gap is unexplained. Their own PointNet++ without TRELLIS gets 52% aneurysm accuracy, while the published PointNet++ on the same dataset gets 87%. A drop that large suggests different preprocessing, training protocol, or split. Without an explanation, the \"vs SOTA\" claim in the abstract is shaky.\n\n3. The dimensionality concern is real. Replacing 3D normals with 1024-d features confounds capacity with transfer. The MLP-only ablation doesn't settle it, because a high-dimensional input can memorize object-level statistics. A control with equally high-dimensional non-transfer features (random Fourier features, or an untrained TRELLIS encoder) would make the cross-domain-transfer claim convincing. The paper's title and conclusion lean on \"cross-domain transfer,\" so this matters.\n\nThe core finding—frozen generative features are surprisingly useful for medical 3D tasks—is likely robust. The exact mechanism is not. That's a good reason to accept it for peer review with major revisions, not to desk reject. The authors should fix the split/overlap, explain the baseline gap, and add a non-transfer high-dimensional control.\n\nWho it's for: people working on 3D medical representation learning, feature transfer, or aneurysm analysis. I'd give it a serious referee pass. Recommendation: send to review; conditional acceptance after the above.","headline":"Frozen TRELLIS features nearly solve aneurysm classification, but the paper's cross-domain transfer claim needs a split clarification and a high-dimensional control.","tokens_in":14483,"tokens_out":2470,"would_cite":true,"duration_ms":26922,"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":"The paper claims that replacing surface normals with TRELLIS's 1024-dimensional latent surface features improves aneurysm classification and segmentation on the Intra3D benchmark and reduces blood-flow simulation error on AnXplore by about","keywords":["intracranial aneurysm","3D point cloud classification","3D mesh segmentation","hemodynamics simulation","feature transfer","TRELLIS","graph neural network","medical imaging"],"falsifier":"Train the same PointNet++, MLP, and GNN pipelines with a matched 1024-dimensional per-point descriptor that carries no pretrained geometric knowledge—for example, a frozen random linear projection of coordinates and normals, or features from an untrained TRELLIS encoder—and compare accuracy and RMSE. If the gains persist, the claim that pretrained generative-model knowledge causes the improvement is refuted; if they vanish, the transfer interpretation is supported.","tokens_in":13663,"feed_emoji":"🧠","tokens_out":9319,"duration_ms":91496,"temperature":0.7,"pith_summary":"This paper sets out to show that latent surface features from TRELLIS, a 3D generative model trained on ordinary non-medical objects, can be transplanted into medical tasks where annotated aneurysm data are scarce. The authors replace the standard point-cloud inputs—3D coordinates plus surface normals—with TRELLIS's 1024-dimensional per-point features, and feed the same features into a graph neural network that predicts blood flow. On the Intra3D benchmark, this lifts classification accuracy on vessels to about 99.9% and on aneurysms to 100% (F1 = 0.999) for PointNet++ at 512 points, and it improves segmentation IoU and Dice scores across the board. On the AnXplore blood-flow benchmark, adding the features cuts the GNN's all-rollout RMSE by roughly 15%. A plain MLP trained on the feature vectors alone, with no spatial coordinates, already separates aneurysms from vessels, which the authors read as evidence that the pretrained encoder has captured reusable shape knowledge that transfers across domains.","feed_headline":"3D generative model features lift aneurysm detection to near-perfect","feed_subtitle":"Swapping normals for TRELLIS's 1024-D surface features raises F1 to 0.999 and cuts blood-flow error by 15%.","key_machinery":"The load-bearing object is the TRELLIS surface feature: a 1024-dimensional latent vector attached to each point of a mesh, produced by rendering the object from multiple views, voxelizing it into a 64x64x64 grid, and encoding the voxel-feature pairs with a transformer-based sparse variational autoencoder that consumes DINOv2 visual features. In the classification and segmentation experiments, these vectors replace surface normals as the point-wise input to PointNet and PointNet++; in the simulation experiments, they are concatenated with the mesh node features of a GNN with masked multi-head self-attention. The feature vector is what carries the cross-domain knowledge, and it is the only var","core_discovery":"The central claim is that the geometric representation learned by a general-purpose 3D generative encoder transfers to a specialized medical domain without any fine-tuning. Concretely, the paper shows that PointNet and PointNet++ fed with TRELLIS surface features instead of point normals achieve near-perfect classification and segmentation on the Intra3D aneurysm benchmark, surpassing the previous baselines from the dataset paper and from 3DMedPT. The same features, concatenated with existing mesh node features in a GNN with transformer blocks, reduce the all-rollout RMSE of blood-flow simulation on AnXplore by about 15%, averaged over the S/1 and L/1 model sizes. The paper also shows that t","pith_inferences":["The paper does not compare against an equally high-dimensional descriptor that carries no learned geometric knowledge, so a reader should treat the cross-domain transfer interpretation as open until a random-feature or untrained-encoder control is run.","The clustering by size and morphology suggests the same features could feed rupture-risk stratification models, but the paper only shows correlation with hemodynamic metrics, not a predictive risk model.","If the transfer effect is real, the strategy may extend to other 3D medical structures with scarce annotations—bone, cardiac, or pulmonary meshes—though only cerebrovascular cases are demonstrated here.","The feature extraction pipeline's use of rendered views means the features may partly encode viewing direction and global object pose; ablating orientation alignment or testing cropped local patches would clarify what the downstream networks actually exploit."],"forward_implications":["Pretrained 3D generative encoders can act as ready-made feature extractors for small medical point-cloud datasets, reducing the need to train 3D encoders from scratch.","With these features, simple architectures reach near-ceiling accuracy on the Intra3D classification benchmark, so further progress on this dataset would have to come from harder or more varied data.","Adding the same fixed features to a hemodynamics GNN improves rollout fidelity by about 15% without changing the solver architecture, suggesting geometry-conditioning is a cheap way to improve physics proxies.","Because the features separate aneurysm and vessel regions in latent space, they could support weakly supervised detection and labeling of vascular structures with far fewer manual annotations.","The one-time encoding cost (about five minutes per object) is amortized across the multiple downstream tasks that reuse the same features."],"supporting_citations":[{"why":"Defines TRELLIS and its encoder, the source of the 1024-dimensional surface features used across all experiments.","marker":"(Xiang et al., 2024)"},{"why":"Provides the Intra3D aneurysm dataset, its classification and segmentation splits, and the baseline results the augmented models are measured against.","marker":"(Yang et al., 2020)"},{"why":"Provides the AnXplore aneurysm models and paired blood-flow simulations used in the GNN experiments.","marker":"(A et al., 2024)"},{"why":"Supplies the GNN-with-transformers architecture for mesh-based hemodynamics that the TRELLIS features are added to.","marker":"(Garnier et al., 2025)"},{"why":"PointNet is one of the two point-cloud backbones whose normal vectors are replaced by TRELLIS features.","marker":"(Qi et al., 2016)"},{"why":"PointNet++ is the other backbone and the model achieving the best reported classification and segmentation results.","marker":"(Qi et al., 2017)"},{"why":"Supplies the 3DMedPT baseline results used in the classification and segmentation comparisons.","marker":"(Yu et al., 2021)"}],"fun_headline_variants":["Generative 3D features sharpen aneurysm detection and flow prediction","TRELLIS features cut aneurysm simulation error by 15%","3D generative features lift aneurysm F1 to 0.999, cut error 15%","Cross-domain 3D features boost aneurysm analysis: near-perfect F1, 15% error cut"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper's load-bearing premise is that the improved results come from geometric knowledge transferred by TRELLIS's encoder, rather than from the sheer dimensionality of the 1024-dimensional feature vectors or from global cues like object orientation and overall shape.","fun_headline_variants_meta":{"raw":{"variants":["Generative 3D features sharpen aneurysm detection and flow prediction","TRELLIS features cut aneurysm simulation error by 15%","3D generative features lift aneurysm F1 to 0.999, cut error 15%","Cross-domain 3D features boost aneurysm analysis: near-perfect F1, 15% error cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001273,"raw_usage":{"total_tokens":5038,"prompt_tokens":731,"completion_tokens":4307,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":4219}},"tokens_in":475,"tokens_out":4307,"duration_ms":31683,"temperature":1.0,"reasoning_tokens":4219,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:06:16.563328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same PointNet++, MLP, and GNN pipelines with a matched 1024-dimensional per-point descriptor that carries no pretrained geometric knowledge—for example, a frozen random linear projection of coordinates and normals, or features from an untrained TRELLIS encoder—and compare accuracy and RMSE. If the gains persist, the claim that pretrained generative-model knowledge causes the improvement is refuted; if they vanish, the transfer interpretation is supported.","supporting_citations":[],"review_version":1}