{"id":"c451eb05-6100-4c9f-ac27-406290235e4d","arxiv_id":"2505.15248","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using real multi-view radiographs from the same study as self-supervised training pairs yields better anatomical representations and downstream veterinary task performance than synthetic single-image augmentations.","lead":"VET-DINO trains a vision transformer on pairs of X-ray images from the same veterinary study, instead of synthetic crops of one image, to learn anatomy that is consistent across views. The method improves downstream veterinary abnormality detection over single-image pretraining and ImageNet-pretrained DINOv2.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-view baseline is not matched on training data: 'one image per study' uses at most 668k images versus 5M for multi-view, so reported gains may reflect data scale, not multi-view pairing.","rationale":"The reader identified study-pair validity as the weakest assumption. I agree that is unvalidated, but I find a more immediate and more load-bearing problem: the single-view baseline is not shown to be matched to the multi-view model on pretraining data quantity or composition. Section 4.1 says the single-view model uses 'only one image per study,' while the multi-view model uses two images per study from a corpus of 5M images across 668k studies. That creates a 7.5x disparity in per-epoch unique images. Table 2's row label 'Single-image Studies' only compounds the ambiguity, suggesting the baseline may even come from a different subset of studies. Without controlling total image count, study count, steps, and epochs, the reported improvements over the single-view baseline do not isolate the effect of multi-view pairing. The anatomical-understanding evidence in Section 4.5 is anecdotal, but the uncontrolled comparison is the primary threat to the paper's central claim. Because a controlled rerun could resolve the issue, I do not move the verdict beyond the reader's CONDITIONAL; the condition should include data-matched baselines. Thus the verdict remains conditional, and my read does not alter the reader's overall conclusion, though it identifies a different primary weakness.","tokens_in":13216,"tokens_out":7011,"duration_ms":61435,"concrete_test":"Retrain Single-view VET-DINO on the full 5M-image pretraining set (or, alternatively, train Multi-view VET-DINO on a matched 668k-image subset) with identical architecture, batch size, and step count, sampling one image per study uniformly in each epoch; then rerun the Section 4.3 k-NN evaluation and Section 4.4 fine-tuning on the same evaluation split. If Multi-view VET-DINO still beats the data-matched single-view baseline by the reported margins, the central claim survives; if the gap closes substantially, the paper's conclusion is not supported by the current experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Sections 4.3 and 4.4 is between Multi-view VET-DINO, which samples crops from two random images within the same study (Section 3.2), and Single-view VET-DINO, defined in Section 4.1 as 'trained on the same dataset but using only one image per study.' The pretraining corpus contains 5 million images from 668,000 studies, so the single-view baseline can see at most 668,000 unique images per epoch—roughly 7.5x fewer training examples unless it actually trains on all images, which contradicts the definition. Table 2 labels the baseline 'Single-image Studies,' suggesting it may have been trained on studies containing only one image rather than on one image sampled per study; either way, the baseline is not matched to Multi-view VET-DINO on data quantity or study composition. Consequently, the large k-NN and fine-tuning gaps (e.g., 0.573 vs 0.280 F1 for Ingesta in the Stomach) cannot be unambiguously attributed to multi-view learning. The paper reports no control for total image count, study count, or number of steps across the two pretraining conditions. Even if all study pairs are anatomically valid, the claim that real multi-view pairs outperform purely synthetic augmentations requires a comparison in which the only difference is the pairing strategy, not the amount of pretraining data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VET-DINO, a self-supervised learning method for veterinary radiographs that extends DINO by sampling crops from two different radiographic views of the same study instead of from a single image. The authors pretrain ViT-S/14 on 5 million canine radiographs from 668,000 studies and evaluate the learned representations through k-NN classification on three gastrointestinal labels and end-to-end fine-tuning on 47 radiographic findings, comparing against a single-view VET-DINO baseline and ImageNet-pretrained DINOv2. The paper also provides qualitative attention-map and patch-embedding-similarity analyses intended to show view-invariant anatomical understanding.","tokens_in":13481,"tokens_out":6039,"duration_ms":47612,"significance":"The idea of using naturally paired medical views as a self-supervised training signal is timely and potentially impactful for medical imaging, where standardized multi-view protocols are common. The scale of the pretraining dataset (5M images) is a notable strength, as are the consistent downstream evaluation protocols across 47 labels. However, the current evidence for the central claim that multi-view pairing outperforms single-view training is weakened by an unmatched baseline and several under-validated assumptions, so the significance as demonstrated is conditional.","major_comments":[{"comment":"The single-view baseline is not matched on data quantity. The text defines it as training on \"only one image per study\" from a corpus of 5M images across 668k studies, which caps the baseline at approximately 668k unique images per epoch, a 7.5x reduction relative to the 5M images used in multi-view training. Table 2 labels the baseline \"Single-image Studies,\" suggesting it may instead have been trained only on studies containing one image, which would be a different subset with its own distribution shift. Either way, the comparison conflates multi-view pairing with total pretraining data, and the large k-NN F1 gaps (e.g., 0.573 vs 0.280 for Ingesta in the Stomach) cannot be unambiguously attributed to multi-view learning. A matched control (e.g., single-view training on all 5M images, or multi-view training subsampled to the same number of images and steps) is needed.","section":"4.1, Table 2"},{"comment":"The k-NN evaluation covers only 3 of the 47 labels, all with very low positive rates (approximately 1-5% in the validation set), and the table reports no confidence intervals. For example, the F1 difference between Multi-view and DINOv2 for Foreign Body in the Small Intestine is 0.3517 vs 0.0738, but with only 733 positive validation examples the variance may be large. This narrow evaluation does not support the abstract's sweeping claim of \"state-of-the-art performance on various veterinary imaging tasks.\"","section":"4.3"},{"comment":"The core assumption that any two randomly selected images within a study constitute valid positive pairs is not validated. The paper does not report the distribution of images per study, whether studies can contain images of different body regions or from different sessions, or any manual inspection of pair quality. If some studies contain mislabeled or merged images, the training signal includes false positives, and the purported mechanism of view-invariant anatomy learning is undermined. The authors should quantify study composition and report a quality audit of sampled pairs.","section":"3.2"},{"comment":"The quantitative evidence for \"anatomical understanding\" is based on 24 image pairs with manually selected anchor patches and reports an average cosine similarity of 0.98 between the anchor and the top-5 most similar patches. This analysis lacks a comparison to the single-view baseline and DINOv2 under the same protocol, lacks a random-pair or same-image control, and does not measure retrieval accuracy of the actual corresponding anatomical location. Without such controls, the high similarity could reflect low-level image statistics rather than view-invariant anatomical correspondence.","section":"4.5.2"},{"comment":"The fine-tuning results are presented without statistical significance or variance estimates. Many improvements over the single-view baseline are modest (e.g., AP for Gastric Distention: 0.778 vs 0.766; Small Kidney: 0.429 vs 0.400), and some labels show Multi-view underperforming one or both baselines (e.g., AP for Gastric Dilatation Volvulus: 0.511 vs 0.541 single-view; Megacolon: 0.010 vs 0.027 DINOv2). The claim of consistent superiority requires reporting confidence intervals or at least the number of runs and standard deviations. Moreover, the claim that VET-DINO achieves \"state-of-the-art performance\" is not supported by comparisons to existing veterinary imaging AI systems or other self-supervised medical imaging methods; the comparisons are limited to two DINO variants, one of which is the authors' own baseline.","section":"4.4, Table 3, Section 2"}],"minor_comments":[{"comment":"The sentence \"A ViT-Tiny model ... is used as the classification layer\" is confusing, as a ViT-Tiny is a transformer encoder rather than a classification layer; please specify the exact head architecture (e.g., linear probe, MLP, or transformer decoder).","section":"4.4"},{"comment":"The column labeled \"Fine-tuning\" appears to contain row labels (\"Multi-image Studies,\" \"Single-image Studies,\" \"None\") rather than fine-tuning settings, making the table hard to parse; additionally, the Multi-view row is missing an entry in that column.","section":"Table 2"},{"comment":"The CNN used to filter images with artifacts and irrelevant views is not described (architecture, training data, or accuracy); please provide details or a reference.","section":"3.3"},{"comment":"The abrupt collapse in the training loss observed in all experiments is not explained; please comment on whether this is an expected DINO training dynamic and how it was handled.","section":"Appendix A, Figure 6"},{"comment":"Several references are incomplete or malformed: [20] lists no title or authors, [4] omits the publication venue, and [24] appears to be a preprint without venue information; please correct these.","section":"References"},{"comment":"The limitation \"evaluation focused primarily on 47 findings\" should say \"evaluation focused on 47 findings,\" since this is the complete label set rather than a subset.","section":"5"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive, and the dataset is substantial. The main obstacle is the unmatched baseline: the single-view comparison suffers from a large data-quantity confound that undermines the central attribution of gains to multi-view pairing. The other major points (pair validity, k-NN breadth, and the anecdotal anatomical-understanding analysis) also need to be addressed for the claims to be convincing. The authors' own limitations section is honest, but it does not mention the baseline matching issue, which is the most serious concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real idea — using the multiple standardized views in a veterinary radiology study as positive pairs for DINO-style self-distillation — and it's executed at impressive scale (5M images, 668k studies, 47 board-certified labels). The multi-view model beats ImageNet DINOv2 on most fine-tuning metrics, which is a meaningful domain-specific result. But the central head-to-head claim against the Single-view VET-DINO baseline is not currently supported. The single-view baseline is defined as one image per study, so it sees at most 668k unique images per epoch, versus 5M for the multi-view model. The paper doesn't match total image count, study count, or training steps across conditions. The differences in Table 2 (e.g., 0.57 vs 0.28 F1) could just be data scale. This is the main thing to fix.\n\nWhat's genuinely good: the adaptation of DINO to real multi-view pairs is natural and well-motivated for medical imaging; the dataset is large and from real clinical workflows; they report both k-NN and fine-tuning; they compare against a strong ImageNet checkpoint; and the limitations section is honest (small architecture, canine only, subjective anatomical evaluation). The attention-map analysis is suggestive, not proof, but they don't oversell it.\n\nOther soft spots, in decreasing order: the k-NN evaluation covers only three gastrointestinal labels with low prevalence; the table formatting hides the fine-tuning column, and there are no error bars anywhere; the cosine-similarity claim rests on 24 manually selected pairs; study grouping quality isn't validated (do all images in a study really share anatomy?); and there's no code or data release. None of these are fatal on their own, but they narrow what a reader can actually verify.\n\nBottom line: this deserves serious peer review — the method and scale are worth engaging — but the current version should not be accepted as-is. The authors need to re-run the single-view baseline with matched data quantity (best: standard DINO on all 5M images, no multi-view pairing), add error bars, and strengthen the anatomical evaluation. If the matched comparison still shows multi-view helps, that's a solid contribution to veterinary SSL and possibly medical imaging more broadly.","headline":"The multi-view idea is sensible and the scale is impressive, but the key single-view baseline sees roughly 7.5x less data, so the paper's central claim currently rests on a confound.","tokens_in":14018,"tokens_out":2788,"would_cite":false,"duration_ms":24988,"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":"VET-DINO claims that drawing training crops from two radiographs of the same patient study, rather than synthetic crops of one image, yields view-invariant anatomical representations that outperform single-view and ImageNet-pretrained…","keywords":["self-supervised learning","multi-view distillation","veterinary radiography","DINO","DINOv2","Vision Transformer","representation learning","canine radiographs"],"falsifier":"Manually inspect a random sample of the 668,000 studies and count how often two images pulled from the same study are not the same patient, same body region, and same imaging session; if that fraction is nontrivial, the training signal contains systematic false positives, and the claimed view-invariance gains should disappear or shrink when those pairs are removed.","tokens_in":13003,"feed_emoji":"🐕","tokens_out":6678,"duration_ms":53917,"temperature":0.7,"pith_summary":"VET-DINO is a self-supervised pretraining method for veterinary radiographs that replaces DINO's synthetic multi-crop augmentation with real multi-view pairing: crops are drawn from two randomly selected images within the same patient study. The paper's central claim is that learning from these real pairs produces view-invariant anatomical representations, including an implied sense of 3D structure from 2D projections, and that this beats learning from synthetic augmentations of single images. On 5 million canine radiographs from 668,000 studies, the multi-view encoder outperforms both a single-view VET-DINO and an ImageNet-pretrained DINOv2 baseline on k-NN F1 for gastrointestinal findings and on fine-tuned AP/ROC AUC for most of 47 radiographic findings. Attention visualizations and patch-embedding cosine similarity are offered as evidence that the model consistently attends to skeletal, soft-tissue, and gastrointestinal structures across views. If the claim holds, self-supervised medical imaging can exploit standardized multi-view protocols rather than adapting natural-image augmentation recipes.","feed_headline":"Real multi-view X-rays beat synthetic crops for veterinary AI","feed_subtitle":"On 5 million canine radiographs, multi-view pretraining outperforms single-view and ImageNet baselines on 47 findings.","key_machinery":"The central object is the real multi-view positive pair: two radiographs randomly selected from the same canine study, used as natural augmentations of one another. The framework keeps DINO's self-distillation machinery, a student ViT-S/14 trained to predict the output of an EMA teacher that sees only global crops, but the student receives global and local crops drawn from both views, so consistency is enforced across projection angles rather than across synthetic distortions of one image. That cross-view consistency is what the paper claims builds view-invariant, quasi-3D anatomical representations.","core_discovery":"The central discovery is that real multi-view pairs within a study are a more powerful self-supervisory signal than augmented views of one image for veterinary radiology. VET-DINO trains a student Vision Transformer to match the representation of an EMA teacher that sees global crops from one randomly chosen radiograph, while the student additionally sees global and local crops from a second radiograph of the same study. The paper argues that this forces the model to become invariant to projection angle and patient positioning while preserving anatomical content, giving it an implied 3D understanding from 2D projections. Empirically, multi-view VET-DINO beats its single-view counterpart and ImageNet-pretrained DINOv2 on the three gastrointestinal k-NN labels and on most of the 47 fine-tuning metrics, with particularly large gains on labels such as Decreased serosal detail, Degenerative Joint Disease, and Uterine Enlargement.","pith_inferences":["Editorial inference: if study grouping is reliable, the same real-pair recipe should transfer to human chest radiography (PA/lateral), mammography views, and dental imaging, where standardized multi-view protocols also exist; the paper only tests canine radiographs.","Editorial inference: the claimed 3D understanding could be quantified directly by cross-view landmark prediction or novel-view synthesis, which the paper lists as future work rather than demonstrating.","Editorial inference: the 0.98 patch-similarity figure comes from only 24 manually selected pairs; a systematic, automated evaluation across many studies would be needed to establish view-invariance as a robust property."],"forward_implications":["Multi-view pretraining raises k-NN F1 for gastrointestinal findings (for example, Ingesta in the Stomach F1 0.573 versus 0.280 for single-view and 0.284 for DINOv2), so frozen VET-DINO features are more useful for label-scarce retrieval.","Fine-tuned detection improves on most of the 47 findings, with the largest gains on labels like Decreased serosal detail (AP 0.700 versus 0.619 and 0.622) and Degenerative Joint Disease (0.589 versus 0.466 and 0.484).","Real study-level pairing works without manual labels or text reports, making the approach suitable for large uncurated clinical archives of radiographs.","Attention maps become concentrated on skeletal, soft-tissue, and gastrointestinal structures that persist across views, supporting the claim that the model learns view-invariant anatomical features.","The same self-supervised recipe could extend to larger ViT architectures and broader datasets, which the paper identifies as a direct next step."],"supporting_citations":[{"why":"Supplies the DINO self-distillation and multi-crop framework that VET-DINO adapts to real multi-view pairs.","marker":"[5]"},{"why":"Provides DINOv2, the ImageNet-pretrained baseline compared against and the source of the pretrained weights.","marker":"[18]"},{"why":"Defines the Vision Transformer architecture (ViT-S/14) used for both teacher and student encoders.","marker":"[10]"},{"why":"Supplies the k-NN evaluation protocol used to measure frozen-feature representation quality.","marker":"[4]"},{"why":"Motivates multi-view radiology by showing lateral chest X-ray views add diagnostic value over frontal views alone.","marker":"[2]"},{"why":"Demonstrates a prior attention-based multi-view classifier for X-rays, a related approach the paper contrasts with its self-supervised pretraining.","marker":"[24]"}],"fun_headline_variants":["Multi-view X-rays teach AI anatomy better than synthetic crops","Real X-ray pairs beat synthetic views for veterinary AI","Canine X-ray multi-views beat single-image tricks for AI","Self-supervised AI learns 3D anatomy from multi-view X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every pair of images from the same study really is two views of the same animal and same anatomy in one imaging session, so that matching crops across them teaches view-invariant anatomy rather than study-specific or patient-specific shortcuts.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view X-rays teach AI anatomy better than synthetic crops","Real X-ray pairs beat synthetic views for veterinary AI","Canine X-ray multi-views beat single-image tricks for AI","Self-supervised AI learns 3D anatomy from multi-view X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2853,"prompt_tokens":910,"completion_tokens":1943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1872}},"tokens_in":526,"tokens_out":1943,"duration_ms":11371,"temperature":1.0,"reasoning_tokens":1872,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:19:49.760783+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Manually inspect a random sample of the 668,000 studies and count how often two images pulled from the same study are not the same patient, same body region, and same imaging session; if that fraction is nontrivial, the training signal contains systematic false positives, and the claimed view-invariance gains should disappear or shrink when those pairs are removed.","supporting_citations":[{"cited_title":"Dinov2: Learning robust visual features without supervi- sion, 2024","cited_arxiv_id":null,"evidence_quote":"Provides DINOv2, the ImageNet-pretrained baseline compared against and the source of the pretrained weights."},{"cited_title":"Do Lateral Views Help Automated Chest X-ray Predictions?","cited_arxiv_id":"1904.08534","evidence_quote":"Motivates multi-view radiology by showing lateral chest X-ray views add diagnostic value over frontal views alone."},{"cited_title":"StudyFormer : Attention-Based and Dynamic Multi View Classifier for X-ray images","cited_arxiv_id":"2302.11840","evidence_quote":"Demonstrates a prior attention-based multi-view classifier for X-rays, a related approach the paper contrasts with its self-supervised pretraining."}],"review_version":1}