Pith. sign in

REVIEW 5 major objections 6 minor 27 references

VET-DINO: Learning Anatomical Understanding Through Multi-View Distillation in Veterinary Imaging

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2505.15248 v1 pith:PHB6KOGP submitted 2025-05-21 cs.CV cs.LG

Andre Dourson , Kylie Taylor , Xiaoli Qiao , Michael Fitzke This is my paper
classification cs.CVcs.LG
keywords self-supervisedlearningmulti-viewdistillationveterinaryradiographyDINODINOv2VisionTransformerrepresentationcanineradiographs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

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.

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 (5)
  1. [4.1, Table 2] 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.
  2. [4.3] 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."
  3. [3.2] 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.
  4. [4.5.2] 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.
  5. [4.4, Table 3, Section 2] 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.
minor comments (6)
  1. [4.4] 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).
  2. [Table 2] 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.
  3. [3.3] 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.
  4. [Appendix A, Figure 6] 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.
  5. [References] 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.
  6. [5] 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.

Circularity Check

2 steps flagged · score 4.0 of 10

The headline mechanism claim is only partially supported: the 'anatomical understanding' evidence measures the model's own cross-view training objective, and the multi-view-vs-single-view comparison is confounded by unmatched pretraining data (5M vs ≤668K images), leaving the causal attribution unseparated from a data-scale input; fine-tuning vs ImageNet DINOv2 retains some independent content…

  1. other [Sections 4.5.1-4.5.2 (Attention Visualization, Patch Embedding Similarity) vs training objective in Section 3.2]
    "To quantify the observed consistency of attention on key anatomical structures, we analyzed the cosine similarity between patch embeddings from different radiographic views of the same study... Across a small sample of 24 image pairs, we calculated an average cosine similarity of 0.98 between the anchor patch and the top 5 most similar patches in the comparison image. This suggests that VET-DINO learns representations that exhibit a degree of view-invariance."

    Section 3.2's training objective directly optimizes cross-view consistency: the student is trained to match the teacher's output across two radiographs of the same study. Section 4.5.2 then measures, on that same trained encoder and on same-study pairs from the training distribution, the cosine similarity between the two views' embeddings, reporting 0.98 (top-5 patches) as evidence of learned 'view-invariance' and anatomical/3D understanding. That metric substantially measures the optimization target itself: cross-view representational agreement on in-distribution same-study pairs is exactly what the loss was fitted to maximize, and the top-5 selection comes from the model's own similarity ranking, so a high value is expected rather than an emergent property demonstrated.

  2. other [Sections 4.1, 3.2-3.3, 4.2, and Table 2 (single-view baseline definition, dataset size, attribution logic)]
    "VET-DINO is trained using crops from two random images within the same study" (3.2); "Single-view VET-DINO: A DINOv2 model trained on the same dataset but using only one image per study" (4.1); "The dataset we used to pretrain our models consists of 5 million X-ray images drawn from approximately 668,000 unique canine studies" (3.3); "By maintaining a consistent training set size across different experimental conditions, we can more confidently attribute performance differences to changes in model architecture or training strategies, rather than variations in the training data itself" (4.2)."

    The paper's central attribution (real multi-view pairing, not data scale, causes the Tables 2-4 gains) requires the only pretraining difference to be the pairing strategy. By the paper's own definitions the inputs differ by construction: with 5M images across 668K studies, 'only one image per study' can present at most 668K unique images per epoch, and Table 2's label 'Single-image Studies' implies an even narrower, never-quantified study subset, while multi-view consumes the full 5M corpus. No control is reported for image count, study composition, or pretraining steps, so the gaps (e.g., 0.5733 vs 0.2795 k-NN F1 for Ingesta) are fully consistent with a data-quantity effect; the claimed 'prediction' co-varies with an uncontrolled input.

full rationale

VET-DINO is an empirical systems paper, so equation-level self-definitional circularity does not apply; I assessed the evidence chain behind the headline claim that 'learning from real multi-view pairs leads to superior anatomical understanding compared to purely synthetic augmentations.' Two partial circularities are identifiable from the paper's own text. (1) The Section 4.5 evidence for 'anatomical understanding' is self-referential: the Section 3.2 distillation loss trains the student to match the teacher across two radiographs of the same study, and Section 4.5.2 then evaluates the trained model via cross-view cosine similarity (0.98, top-5 patches) and calls it learned view-invariance — a metric that substantially coincides with the optimized objective; the attention-map evidence likewise interprets the model's own saliency under manually assigned anatomical labels. The paper's limitation statement concedes only the manual-selection issue. (2) The central single-view baseline is not input-matched: with 5M images from 668K studies, 'only one image per study' caps unique pretraining images per epoch at 668K (Table 2 even labels it 'Single-image Studies'), while multi-view consumes the full corpus, so the reported k-NN and fine-tuning gaps are consistent with a data-scale effect and the causal attribution is not identified; the paper's own Section 4.2 attribution-logic sentence is violated by this comparison. Offset factors: the comparison against the externally pretrained ImageNet DINOv2 is independently falsifiable and favors VET-DINO despite DINOv2 having far more pretraining data; self-citations ([11], [24]) are background and not load-bearing; no uniqueness theorem or ansatz is smuggled via citation; the domain adaptation to veterinary multi-view studies is a genuine application, not a renaming. I also note the abstract promises 'view synthesis' experiments absent from the full text, a manuscript-consistency defect rather than circularity. Overall, the central claims have independent content but the mechanism-specific evidence is partially self-referential and confounded.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or mathematical entities are invented. The free parameters are standard SSL hyperparameters, and the domain assumptions are explicit. The main load-bearing assumption is the validity of study-level pairing as a source of view-invariant positive pairs.

free parameters (3)
  • k in k-NN classifier = 3
    Selected by grid search with 5-fold cross-validation to maximize F1 (Section 4.1). This is a hyperparameter fitted to the evaluation data, though its impact on the central multi-view vs single-view comparison is likely modest.
  • Multi-crop window sizes and number of crops = 2 global (224x224) + 10 local (98x98) per study
    Copied from DINO with minor modifications; chosen by hand based on prior art, not fitted to veterinary data. Affects training signal but is a standard SSL hyperparameter.
  • Projection layer dimension = 16384
    Reduced from DINOv2's 65536 to save memory; authors claim minimal impact on downstream performance (Section 3.4, 4.4). This is a design choice, not a fitted parameter in the classical sense.
assumptions (3)
  • domain assumption Images within a canine study are views of the same patient's anatomy acquired under standardized protocols.
    This is the core assumption enabling multi-view training. It enters in Section 3.2 where two random images from the same study are used as positive pairs. If study grouping is noisy (e.g., mixed sessions or body regions), the training signal is corrupted.
  • domain assumption Board-certified radiologist labels on 47 findings are accurate and consistently applied across the evaluation dataset.
    The evaluation depends entirely on these manual labels (Section 4.2). The paper does not report inter-observer agreement or label quality checks, so label noise could affect the measured performance differences.
  • standard math The DINO self-distillation framework, including EMA teacher and centering, is a standard and valid SSL method.
    The paper builds on DINO/DINOv2 and inherits their design without re-deriving it. This is a standard background assumption in modern SSL.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VET-DINO: Learning Anatomical Understanding Through Multi-View Distillation in Veterinary Imaging." pith.science (2026). https://pith.science/paper/PHB6KOGP

@misc{pith2026250515248,
  author       = {Pith},
  title        = {Pith review of: VET-DINO: Learning Anatomical Understanding Through Multi-View Distillation in Veterinary Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PHB6KOGP}},
  note         = {Machine review of arXiv:2505.15248}
}
read the original abstract

Self-supervised learning has emerged as a powerful paradigm for training deep neural networks, particularly in medical imaging where labeled data is scarce. While current approaches typically rely on synthetic augmentations of single images, we propose VET-DINO, a framework that leverages a unique characteristic of medical imaging: the availability of multiple standardized views from the same study. Using a series of clinical veterinary radiographs from the same patient study, we enable models to learn view-invariant anatomical structures and develop an implied 3D understanding from 2D projections. We demonstrate our approach on a dataset of 5 million veterinary radiographs from 668,000 canine studies. Through extensive experimentation, including view synthesis and downstream task performance, we show that learning from real multi-view pairs leads to superior anatomical understanding compared to purely synthetic augmentations. VET-DINO achieves state-of-the-art performance on various veterinary imaging tasks. Our work establishes a new paradigm for self-supervised learning in medical imaging that leverages domain-specific properties rather than merely adapting natural image techniques.

Figures

Figures reproduced from arXiv: 2505.15248 by the authors.

Figure 1
Figure 1. Multi-view VET-DINO architecture. Two radiographic views are randomly selected from a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Samples of radiographs from the validation dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The above figure displays ventrodorsal and left lateral radiographic projections of a canine [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparative analysis of the Multi-view VET-DINO model against a Single-view VET-DINO [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualization of cosine similarity between patch embeddings for a multi-view VET-DINO [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: VET-DINO pre-training loss illustrates the training loss trajectory during the model’s self [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Single-view VET-DINO Architecture. Unlike Multi-view VET-DINO, a single radiographic [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Patch-level similarity is assessed by comparing embeddings extracted from the final layer [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

  1. [1]

    Maira-2: Grounded radiology report generation.arXiv preprint arXiv:2406.04449, 2024

    Shruthi Bannur, Kenza Bouzid, Daniel C Castro, Anton Schwaighofer, Anja Thieme, Sam Bond- Taylor, Maximilian Ilse, Fernando P´ erez-Garc ´ ıa, Valentina Salvatelli, Harshita Sharma, et al. Maira-2: Grounded radiology report generation.arXiv preprint arXiv:2406.04449, 2024

    arXiv 2024

  2. [2]

    Do Lateral Views Help Automated Chest X-ray Predictions?

    Hadrien Bertrand, Mohammad Hashir, and Joseph Paul Cohen. Do lateral views help automated chest x-ray predictions?arXiv preprint arXiv:1904.08534, 2019

    work page Pith review arXiv 1904

  3. [3]

    Artificial intelligence in veterinary diagnostic imaging: Perspectives and limi- tations.Research in Veterinary Science, page 105317, 2024

    Silvia Burti, Tommaso Banzato, Simon Coghlan, Marek Wodziniski, Margherita Bendazzoli, and Alessandro Zotti. Artificial intelligence in veterinary diagnostic imaging: Perspectives and limi- tations.Research in Veterinary Science, page 105317, 2024

    work page 2024

  4. [4]

    Unsupervised learning of visual features by contrasting cluster assignments.CoRR, abs/2006.09882, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.CoRR, abs/2006.09882, 2020

    arXiv 2006

  5. [5]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´ e J´ egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  6. [6]

    Med3d: Transfer learning for 3d medical image analysis

    Sihong Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis. arXiv preprint arXiv:1904.00625, 2019

    arXiv 1904

  7. [7]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PMLR, 2020

    2020

  8. [8]

    Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V

    Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical auto- mated data augmentation with a reduced search space, 2019

    work page 2019

Show all 27 references
  1. [9]

    Viewclr: Learning self-supervised video representation for unseen viewpoints

    Srijan Das and Michael S Ryoo. Viewclr: Learning self-supervised video representation for unseen viewpoints. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5573–5583, 2023

    2023

  2. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  3. [11]

    Michael Fitzke, Conrad Stack, Andre Dourson, Rodrigo M. B. Santana, Diane Wilson, Lisa Ziemer, Arjun Soin, Matthew P. Lungren, Paul Fisher, and Mark Parkinson. Rapidread: Global deployment of state-of-the-art radiology ai for a large veterinary teleradiology practice, 2021

    2021

  4. [12]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  5. [13]

    Artificial intelligence in veterinary diagnostic imaging: A literature review.Veterinary Radiology & Ultrasound, 63:851– 870, 2022

    Erin Hennessey, Matthew DiFazio, Ryan Hennessey, and Nicky Cassel. Artificial intelligence in veterinary diagnostic imaging: A literature review.Veterinary Radiology & Ultrasound, 63:851– 870, 2022

    2022

  6. [14]

    Imagenet classification with deep con- volutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep con- volutional neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012

    2012

  7. [15]

    Sgdr: Stochastic gradient descent with warm restarts, 2017

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts, 2017

    2017

  8. [16]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

    2019

  9. [17]

    Comparison of two-vs

    Christopher P Ober and Don Barber. Comparison of two-vs. three-view thoracic radiographic studies on conspicuity of structured interstitial patterns in dogs.Veterinary Radiology & Ultra- sound, 47(6):542–545, 2006. 12

    2006

  10. [18]

    Dinov2: Learning robust visual features without supervi- sion, 2024

    Maxime Oquab, Timoth´ ee Darcet, Th´ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nico- las Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael...

    2024

  11. [19]

    Rad-dino: Exploring scalable medical image encoders beyond text supervision.arXiv preprint arXiv:2401.10815, 2024

    Fernando P´ erez-Garc ´ ıa, Harshita Sharma, Sam Bond-Taylor, Kenza Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Rad-dino: Exploring scalable medical image encoders beyond text supervision.arXiv p...

    2024 arXiv

  12. [20]

    Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning

    P Rajpurkar. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. ArXiv abs/1711, 5225, 2017

    2017

  13. [21]

    Facenet: A unified embedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 815–823, 2015

    2015

  14. [22]

    Mv2mae: Multi-view video masked autoencoders.arXiv preprint arXiv:2401.15900, 2024

    Ketul Shah, Robert Crandall, Jie Xu, Peng Zhou, Marian George, Mayank Bansal, and Rama Chellappa. Mv2mae: Multi-view video masked autoencoders.arXiv preprint arXiv:2401.15900, 2024

    2024 arXiv

  15. [23]

    Medaug: Contrastive learning leveraging patient metadata improves representations for chest x-ray interpretation

    Yen Nhi Truong Vu, Richard Wang, Niranjan Balachandar, Can Liu, Andrew Y Ng, and Pranav Rajpurkar. Medaug: Contrastive learning leveraging patient metadata improves representations for chest x-ray interpretation. InMachine Learning for Healthcare Conference, pages 755–769. PMLR, 2021

    2021

  16. [24]

    Studyformer: Attention-based and dynamic multi view classifier for x-ray images.arXiv preprint arXiv:2302.11840, 2023

    Lucas Wannenmacher, Michael Fitzke, Diane Wilson, and Andre Dourson. Studyformer: Attention-based and dynamic multi view classifier for x-ray images.arXiv preprint arXiv:2302.11840, 2023

    2023 arXiv

  17. [25]

    Pytorch image models.https://github.com/rwightman/ pytorch-image-models, 2019

    Ross Wightman. Pytorch image models.https://github.com/rwightman/ pytorch-image-models, 2019

    2019

  18. [26]

    The role of artificial intelligence in clinical imaging and workflows.Veterinary Radiology & Ultrasound, 63:897–902, 2022

    Diane U Wilson, Michael Q Bailey, and John Craig. The role of artificial intelligence in clinical imaging and workflows.Veterinary Radiology & Ultrasound, 63:897–902, 2022

    2022

  19. [27]

    Preservational learning improves self-supervised medical image models by reconstructing diverse contexts

    Hong-Yu Zhou, Chixiang Lu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. Preservational learning improves self-supervised medical image models by reconstructing diverse contexts. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3499–3509, 2021. 13 A...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.