{"id":"e5f832b1-9fbb-4889-87d3-ac0ce2f14f26","arxiv_id":"2507.08254","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A train-free method that compresses 3D medical volumes into small embeddings via a frozen 2D foundation model and random projections, outperforming several medical-volume pretrained models on benchmark tasks.","lead":"Raptor embeds 3D medical scans by running a frozen off-the-shelf 2D image model on the scan's slices in three directions, then shrinking the resulting features with random projections. It needs no training on 3D data and, on ten medical imaging tasks, matches or beats models that were pretrained on large collections of medical volumes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Regression SOTA may be an artifact of head asymmetry: Raptor uses a tuned 3-layer MLP while baselines get a single linear layer; a head-matched rerun is needed.","rationale":"The reader's weakest_assumption (Appendix A.6 alignment) is real and the paper itself documents violations (Fracture axial view, alpha=0 peaks), but it mainly invalidates a supporting theoretical guarantee; the empirical SOTA claim could stand even if the bound fails. The more decisive threat to the paper's headline is that the regression comparison, one of the ten tasks and the source of the largest reported gains, is not head-matched. Section 4.1 gives baselines a single linear layer while Raptor gets a 3-layer MLP; Table 5 then reports large r2 advantages. This is a testable confound: if the same MLP were applied to baseline latents, the margins might shrink. The paper deserves credit for public code, frozen DINOv2 features, and reproducible random projections, so the check is straightforward and should settle the issue. The classification results, which use linear heads on both sides, are less affected. I therefore retain the reader's CONDITIONAL verdict pending the head-matched regression rerun; no verdict change.","tokens_in":23559,"tokens_out":7417,"duration_ms":82658,"concrete_test":"Re-run the UKBB regression benchmark with the downstream head held fixed across methods: (a) train the exact 3-layer MLP from Appendix A.2 (same hidden width 256, BatchNorm, ReLU, MSE loss, 50 epochs, validation early stopping) on the SuPreM, Merlin, MISFM, SLIViT, and VoCo latents using the same 60/20/20 splits; and (b) also train a single linear head on Raptor embeddings. Report per-region r2 and averages for both protocols. If the current Raptor margins over SuPreM/Merlin (Table 5 averages 0.389 vs 0.299/0.313) shrink to within noise or reverse under (a), the headline regression SOTA is a head-capacity artifact; if margins persist under both (a) and (b), the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 fixes the downstream protocol: baselines receive a single linear layer above their latent space, while Raptor embeddings are fed to logistic regression for classification and to an MLP 'tuned up to 3 layers' for regression (Appendix A.2). Table 5 and the Discussion's '+24% over Merlin, +30% over SuPreM' therefore compare different predictor classes, not just different embeddings. The UKBB regression benchmark is one of the ten headline tasks and the source of the largest relative gains. A 3-layer MLP can exploit nonlinear structure and interactions that a linear head cannot; with roughly 885 training volumes and a 77,100-dimensional Raptor input, the MLP's extra capacity (with BatchNorm, ReLU, and early stopping) is a plausible confound. Parameter count alone does not resolve this: the linear heads on SuPreM/Merlin latents are also large, but they are constrained to be linear. The classification results are less affected because logistic regression is itself linear, but the regression claim is not a fair test of embedding quality as stated. The paper does not report head-matched comparisons or error bars, so the central 'superior performance' claim is conditional on an untested assumption that the same head class would preserve the margins.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Raptor (Random Planar Tensor Reduction) is a train-free method for embedding 3D medical volumes. It applies a frozen 2D DINOv2-L encoder to all slices in three orthogonal axes, mean-pools the slice-level token embeddings per axis, compresses the resulting 3 x d x p^2 tensor with random projections to dimension K per patch, and flattens the result into a 3Kp^2 vector. The authors evaluate Raptor on six 3D MedMNIST classification tasks, three additional classification datasets (CC-CCII, CTRG-C, CTRG-B), and a UK Biobank brain MRI regression benchmark with ten groups of imaging-derived phenotypes. They compare against 3D ResNet, MAE-finetuned ViT, and several medical-volume pretraining methods (SuPreM, MISFM, Merlin, VoCo, SLIViT), reporting that Raptor achieves the highest average classification and regression scores while using embeddings far smaller than raw volumes and smaller than most competing latent spaces. The paper also provides ablation studies on the number of random projections, the number of viewpoints, and a controlled simulation of spatial resolution limits, plus a formal analysis of distance preservation in Appendices A.5 and A.6.","tokens_in":23815,"tokens_out":4717,"duration_ms":50601,"significance":"The core idea is appealing and practically valuable: a trained 3D model can be bypassed entirely by reusing a frozen 2D foundation model plus random projections, which is genuinely train-free for embedding generation and scales sub-cubically with volume size. The classification results are credible and broadly consistent with the claim that Raptor is competitive with or better than several medical-volume foundation models on the benchmarks tested. The paper is also honest in reporting the failure modes: the size-detection simulation shows AUC near 0.5 at 16 px, and the Appendix A.6 analysis explicitly notes the alpha_j > 0 assumption is violated in Fracture3D. If the regression comparison is made head-matched and the reported aggregate numbers are corrected, the contribution would be a solid, reproducible baseline for resource-constrained volumetric analysis. The main weakness is that the paper's headline claims overstate the measured margins and rest in part on an unfair predictor-class comparison for the regression benchmark.","major_comments":[{"comment":"The regression comparison is not head-matched: all baselines receive a single linear layer above their latent space, while Raptor embeddings are fed to an MLP tuned up to 3 layers (Appendix A.2). Since the UKBB regression benchmark produces the largest relative gains (+24% over Merlin, +30% over SuPreM), these margins may reflect predictor capacity rather than embedding quality. Please rerun the regression benchmark with a common predictor class (for example, ridge regression or the same MLP architecture) for all embeddings, and report both configurations, or restrict the regression claims accordingly.","section":"Section 4.1 and Table 5"},{"comment":"The numerical claims are inconsistent across the paper. The abstract states +3% over SuPreM and +6% over MISFM, while Section 4.2 reports average classification improvements of +2% over SuPreM and +4% over MISFM, and the Discussion states a '3% accuracy gain over the next best approach (SuPreM)'. Please specify whether these are AUROC or accuracy, whether they include all ten tasks or only the classification tasks, and reconcile the numbers in all three places.","section":"Abstract, Section 4.2, Section 6"},{"comment":"The lower bound d_Raptor >= (1-epsilon)/D * alpha_min * d_raw requires alpha_j > 0 for every slice, but the paper itself reports negative alpha_j in the axial view of Fracture3D and Figure A.3 shows peaks at alpha = 0 due to empty slices. Under these violations the lower bound does not hold, so the claim that 'cluster centers remain Omega(beta)-separated, thereby guaranteeing class separability' is not established for those datasets. Please state the guarantee as explicitly conditional and report the alpha_j diagnostics for all benchmark datasets, not only a subset.","section":"Appendix A.6 and Figure A.3"},{"comment":"The main results report only a single split per dataset without repeated runs or confidence intervals. The headline margins over SuPreM are small (about 1-2 AUROC points on average) and Raptor does not win on every dataset, so the 'superior performance' claim would be more convincing with variance estimates, at least for the closest competitor and on the datasets where the margins are within a few points.","section":"Tables 3, 4, and 5"}],"minor_comments":[{"comment":"The notation for the partial sum is confusing: S_j is introduced as S_j = sum_{k=1}^j Delta_k but written as 'Pk=j k=1'; please clarify the indexing and align the alpha_j assumption with the induction step.","section":"Appendix A.6"},{"comment":"The class percentages for CTRG-C sum to more than 100% and the dataset is described as multi-label, while the paper reports AUROC and ACC in the MedMNIST style; please clarify how accuracy is computed for multi-label tasks.","section":"Table A.5"},{"comment":"The method name is typeset inconsistently as 'V oCo' with an extra space in several places; please fix the typography.","section":"Table 1"},{"comment":"The size-detection simulation shows AUC ~0.5 at 16px, which is an honest limitation, but the abstract and Discussion do not mention this resolution limit; consider adding one sentence to the Discussion to balance the scalability claims.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful, simple idea and a fairly thorough empirical study, but the abstract overstates the results compared to Section 4.2, and the regression SOTA claim rests on a head asymmetry that should be corrected before publication. The code link is a positive sign, but I did not verify the repository. The theoretical appendix is honest about its assumptions, yet the 'guaranteeing class separability' sentence goes beyond what the analysis supports; tightening that language and adding head-matched regression results would make the contribution solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Raptor is worth reading: the combination of a frozen DINOv2-L, triplanar slice mean-pooling, and per-patch random projections is new to me, it is genuinely train-free, and the code is out. The embedding is around 77k floats (K=100) versus raw volumes of 256^3, so the compression story is real. On classification, Raptor edges out SuPreM and the other medical-pretrained baselines on average—about +1.5 AUROC and +0.7 accuracy across the nine tasks—but the margin is thin and it loses on a few datasets. The abstract's +3% over SuPreM, +6% over MISFM do not match Section 4.2's +2% and +4%, nor the tables. The internal numbers need a pass.\n\nThe larger concern is the regression benchmark. Baselines get a single linear head; Raptor gets a validation-tuned 3-layer MLP. On UKBB—885 training volumes, 162 targets, 77k inputs—that extra capacity is a plausible confound. The +24% over Merlin and +30% over SuPreM are exactly the largest relative wins and exactly where the predictor classes differ. Without a head-matched rerun, I'd not treat the regression claim as established. The classification comparisons are fairer since logistic regression is linear, but even there, a few wins are within a percentage point.\n\nCredit where it's due: the paper publishes its weak spot. Appendix A.6 states the alpha_j > 0 alignment assumption behind mean-pooling, shows it is violated in Fracture3D, and Section 5.3's size-detection experiment demonstrates the expected loss of small-feature information. That is honest, and the ablations on K, views, and encoders are useful. Reproducibility is solid: code, public datasets, standard splits.\n\nShould it go to review? Yes. The method is simple, useful, and the flaws are fixable. I'd ask a referee to demand: (1) matched predictor classes for regression, (2) one consistent set of average-improvement numbers across abstract, body, and discussion, and (3) error bars or seed variance on the main result tables. After that, the paper should be much cleaner. For a reading group, I'd bring it anyway to talk about evaluation protocols; but I'd cite it with a note about the head asymmetry.","headline":"Raptor is a genuinely train-free and useful 3D embedding recipe, but the headline regression wins look like an MLP-vs-linear-head artifact; the classification gains are real but thin.","tokens_in":24385,"tokens_out":6300,"would_cite":true,"duration_ms":67390,"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":"Raptor shows that a frozen 2D vision model, applied to orthogonal slices and compressed by random projections, produces train-free embeddings that beat medical-volume pretrained models on ten benchmarks.","keywords":["3D medical volumes","train-free embeddings","random projections","DINOv2","volumetric classification","volumetric regression","dimensionality reduction","foundation models"],"falsifier":"Measure Raptor's AUC on a dataset whose only signal is a small, localized abnormality (a 10–16px sphere or fracture line): the paper's size simulation predicts chance-level AUC ~0.5 at that scale, so a real clinical dataset where human readers reliably detect such lesions while Raptor stays at chance would settle whether the embeddings preserve small-feature information.","tokens_in":23327,"feed_emoji":"🧠","tokens_out":7534,"duration_ms":78115,"temperature":0.7,"pith_summary":"This paper sets out to show that a frozen 2D image foundation model, never trained on medical data, can serve as a general-purpose feature extractor for 3D medical volumes if its slice-wise outputs are averaged and compressed properly. The proposed method, Raptor, applies DINOv2-L to every cross-section of a volume along three orthogonal axes, mean-pools the slice features per axis, and reduces each with random projections, producing an embedding of $3Kp^2$ numbers with no training. On ten classification and regression benchmarks the paper reports state-of-the-art scores, beating medical-volume pretrained models by 3–14% in accuracy while using embeddings up to ~99% smaller than the raw voxels and smaller than competing latents. A sympathetic reader would care because it suggests that expensive 3D pretraining may be avoidable for many volumetric tasks, widening access to high-dimensional medical analysis.","feed_headline":"Frozen 2D model beats 3D medical models without any training","feed_subtitle":"Raptor shrinks a 256³ MRI to ~77k numbers without training, then tops ten medical benchmarks.","key_machinery":"The load-bearing object is the Raptor embedding $\\Phi(x)=\\operatorname{concat}_{i=1}^3 R\\,\\frac{1}{D}\\sum_{j=1}^D \\phi(s_{ij})$, where $\\phi$ is the frozen DINOv2-L vision transformer (a transformer that encodes images as grids of patch tokens), $s_{ij}$ are slices in the axial, coronal, and sagittal directions, and $R\\in\\mathbb{R}^{K\\times d}$ is a Gaussian random projection matrix. The mean-pooling aggregates the volume dimension while the random projection compresses the token dimension, and the Johnson–Lindenstrauss lemma supplies the pairwise-distance guarantee. The whole pipeline is linear after the encoder, which also makes the runtime $O(p^2 d N (D+K))$ and sub-cubic in volume size.","core_discovery":"The central claim is that distance structure in DINOv2-L slice embeddings survives both averaging and random projection, so the resulting volume descriptor preserves enough semantic and spatial information to outperform models pretrained on medical volumes. Concretely, for two volumes the Raptor distance $d_{\\text{Raptor}}$ is shown to remain between a constant times the raw embedding distance and a constant times $\\sqrt{D}\\,d_{\\text{raw}}$, provided slice-level embedding differences are aligned across slices ($\\alpha_j>0$). The paper reports that this holds empirically on most of its ten tasks, where Raptor reaches the best AUROC/accuracy in six of nine classification datasets and the best $r^2$ in nine of ten regression regions, on embeddings of size $3\\times 100\\times 16\\times 16$ (or ten projections for Raptor-B).","pith_inferences":["[Editorial] The same 'orthogonal slices + frozen 2D encoder + random projection' recipe should transfer to non-medical volumetric data, such as microscopy volumes or 3D objects, since nothing in the pipeline is modality-specific.","[Editorial] The size-detection simulation suggests a sharp spatial-resolution floor: features much smaller than a 16px ViT patch are likely to be averaged away, so tasks requiring detection of tiny lesions or thin connected structures would need patch-level or multi-scale pooling rather than whole-slice averaging.","[Editorial] A testable prediction is that Raptor embeddings with different random seeds at K=100 are nearly interchangeable (the paper reports standard deviation below 0.001 in AUC), implying downstream classifiers should be stable across seeds even when the underlying 2D encoder is changed.","[Editorial] The paper's own observation of negative $\\alpha_j$ in Fracture3D predicts that datasets with abrupt slice-to-slice changes, such as volumes containing multiple disconnected organs, will degrade toward chance, and fixing this may require a robust aggregation other than mean-pooling."],"forward_implications":["A 256³ volume becomes a 76,800-dimensional vector (K=100) or 7,680-dimensional vector (K=10) with no training, which fits on-chip for downstream logistic regression or small MLPs.","Raptor-B at K=10 matches SuPreM's average classification accuracy while being 28.8× smaller in embedding size, so the compression itself does not explain the accuracy gap.","Because the encoder is frozen and model-agnostic, replacing DINOv2-L with a stronger 2D foundation model should transfer directly to volumes without re-architecting anything.","The ~6.5s per 256³ volume on a single consumer GPU, with no training run, makes large-scale volume analysis feasible in data-scarce laboratories.","The theory ties class separability in the original embedding space to an $\\Omega(\\beta)$ separation guarantee after compression, so Raptor's downstream behavior is not purely empirical."],"supporting_citations":[{"why":"Supplies the frozen DINOv2-L encoder whose slice tokens are pooled and projected; it is the feature backbone Raptor relies on.","marker":"Oquab et al., 2023"},{"why":"Provides the elementary Johnson–Lindenstrauss proof that underpins the error-preservation guarantee for random projections.","marker":"Dasgupta & Gupta, 2003"},{"why":"Provides the 3D MedMNIST benchmark, its predetermined splits, and the AUROC/accuracy evaluation convention used for six of the classification datasets.","marker":"Yang et al., 2023"},{"why":"Supplies SuPreM, the strongest medical-volume pretrained baseline that Raptor outperforms and the reference point for the 3% accuracy gain.","marker":"Li et al., 2024"},{"why":"Supplies SLIViT, a 2D-pretrained 3D model that serves as a baseline Raptor exceeds on the benchmarked tasks.","marker":"Avram et al., 2024"},{"why":"Supplies Merlin, a chest-CT pretrained baseline whose regression performance Raptor compares against on the brain MRI tasks.","marker":"Blankemeier et al., 2024"},{"why":"Supplies MISFM, a medical-image segmentation foundation model used as a pretrained baseline in all benchmarks.","marker":"Wang et al., 2023"},{"why":"Supplies VoCo, a 160K-CT pretrained model that represents the largest-scale medical-volume baseline Raptor is compared with.","marker":"Wu et al., 2024a"},{"why":"Supplies the brain MRI volumes and imaging-derived phenotype targets used for the ten regression benchmarks.","marker":"Bycroft et al., 2018"},{"why":"Motivates the choice of random projection for scalable dimensionality reduction with distance-preservation properties.","marker":"Xie et al., 2017"}],"fun_headline_variants":["Frozen 2D model slices beat trained 3D medical models","Train-free Raptor tops ten medical volume benchmarks","Random projections preserve slices, outdo trained 3D nets","Zero-training embedding wins on medical volumes with 2D model","DINOv2 slices plus projections: new top for medical volumes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that averaging features over all slices in each of three directions does not cancel the signal distinguishing two volumes: the formal guarantee requires slice-level embedding differences to point in roughly the same direction ($\\alpha_j>0$), and the paper's own results show this fails on Fracture3D's axial view and on empty slices, where averaged differences can vanish.","fun_headline_variants_meta":{"raw":{"variants":["Frozen 2D model slices beat trained 3D medical models","Train-free Raptor tops ten medical volume benchmarks","Random projections preserve slices, outdo trained 3D nets","Zero-training embedding wins on medical volumes with 2D model","DINOv2 slices plus projections: new top for medical volumes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":1187,"prompt_tokens":928,"completion_tokens":259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":172}},"tokens_in":544,"tokens_out":259,"duration_ms":3777,"temperature":1.0,"reasoning_tokens":172,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:23:31.393390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure Raptor's AUC on a dataset whose only signal is a small, localized abnormality (a 10–16px sphere or fracture line): the paper's size simulation predicts chance-level AUC ~0.5 at that scale, so a real clinical dataset where human readers reliably detect such lesions while Raptor stays at chance would settle whether the embeddings preserve small-feature information.","supporting_citations":[],"review_version":1}