Pith. sign in

REVIEW 5 major objections 6 minor 9 cited by

Curia: A Multi-Modal Foundation Model for Radiology

T0 review · 5 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Self-supervised pre-training on 200 million unlabeled CT and MRI slices from one hospital yields a radiology foundation model that matches or exceeds resident radiologists and prior foundation models across a 19-task benchmark.

desk verdict Serious radiology foundation model with a strong benchmark and a genuinely interesting cross-modal transfer result, but the 'beats residents' headline needs better support before I would repeat it. read the letter →

arxiv 2509.06830 v1 pith:2PP5LQZI submitted 2025-09-08 cs.CV cs.LG

Corentin Dancette , Julien Khlaut , Antoine Saporta , Helene Philippe , Elodie Ferreres , Baptiste Callard , Théo Danielou , Léo Alberge
show 15 more authors
Léo Machado Daniel Tordjman Julie Dupuis Korentin Le Floch Jean Du Terrail Mariam Moshiri Laurent Dercle Tom Boeken Jules Gregory Maxime Ronot François Legou Pascal Roux Marc Sapoval Pierre Manceron Paul Hérent
This is my paper
classification cs.CVcs.LG
keywords foundationmodelradiologyself-supervisedlearningCTMRIcross-modalgeneralizationfew-shotCuriaBench
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

The paper claims that a single self-supervised model trained on unlabeled routine CT and MRI scans from one hospital can serve as a general radiology backbone, replacing the usual one-task-one-model approach. That claim is supported by a new 19-task benchmark on which Curia, without fine-tuning its backbone, matches or beats two existing medical foundation models and, on most tasks, final-year resident radiologists. The authors identify two emergent properties that would matter clinically if confirmed: features transfer across CT and MRI without paired data, and near-maximal performance can be reached from very few labeled examples. They also show that frozen imaging features predict kidney-cancer survival better than the standard local tumor stage. If true, the result implies that large unlabeled hospital archives, not curated annotated datasets, are the main ingredient for broad radiology AI.

What carries the argument

DINOv2 self-supervised pre-training on a vision transformer: a teacher–student objective that combines class-token feature alignment, patch-level masked prediction, and auxiliary regularization losses, trained with strong cropping augmentations on 200M unlabeled CT/MRI slices. This is the mechanism that produces the frozen features reused by all downstream heads; the paper argues that the scale and routine-clinical character of the corpus, not task supervision, is what gives the features their generality and their CT-MRI alignment.

What would settle it

Evaluate Curia-L with the paper's frozen-backbone linear-probe protocol on CT examinations collected from two or three hospitals with different scanner vendors and patient populations, using the same 54-class organ-recognition task. If accuracy falls toward the from-scratch ViT level or the CT-to-MRI transfer drop grows well beyond the reported 9.17 percentage points, the single-center pre-training claim of broad generalizability would be refuted.

Watch

Extended reading notes

Core claim

Curia is a pair of vision transformers (86M and 300M parameters) pre-trained with the DINOv2 algorithm on roughly 200 million 2D CT and MRI images drawn from the entire multi-year imaging output of a single hospital. The authors' central claim is that this unlabeled, single-center corpus is sufficient to learn a transferable representation of radiological anatomy and pathology. They evaluate the model by freezing the backbone and training only lightweight prediction heads — linear, cross-attention, regression, Cox survival, and SAM-based segmentation heads — on 19 public tasks spanning organ recognition, oncology, musculoskeletal disease, emergencies, neurodegeneration, and infection. Curia

Load-bearing premise

The model's pre-training corpus is the imaging output of a single hospital, and the paper assumes that one center's mix of scanners, protocols, and patients is varied enough to represent radiology as a whole; if it is not, the benchmark results will overstate how well Curia generalizes outside that center.

Editorial extensions

If this is right

  • A single radiology foundation model can replace many task-specific models: Curia reaches leading results on 19 tasks spanning organ recognition, oncology, trauma, infection, musculoskeletal, and neurodegenerative imaging using only lightweight heads on frozen features.
  • Low-data regimes become practical: near-maximal anatomical accuracy from tens of labeled examples per class, and useful malignancy AUC from 50 or more examples, so annotation-hungry tasks can be bootstrapped quickly.
  • Cross-modality transfer is usable: a linear head trained on CT organ recognition transfers to MRI with only a 9-point drop, and CT-to-MRI registration improves over existing foundation models, so labels in one modality may help another.
  • Imaging features carry prognostic signal: kidney-cancer survival prediction from baseline CT features exceeds the local T-stage baseline, suggesting imaging-derived biomarkers can complement current staging.
  • Frozen-backbone evaluation is sufficient to stage progress: because Curia is compared without fine-tuning the backbone, the benchmark results reflect the quality of the learned representation itself.

Reading between the lines

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

  • The paper's single-center pretraining makes vendor and protocol shift the main untested risk; a direct follow-up is to pre-train on a second hospital's archive and measure whether the CuriaBench gains shrink or grow.
  • The reported cross-modal transfer implies a practical bootstrapping recipe the authors do not test: use CT-trained heads to auto-label MRI volumes, then train MRI-specific heads with those pseudo-labels to reduce annotation cost.
  • Because all benchmarks use public datasets with different preprocessing, holding head architecture and preprocessing fixed across models would isolate how much of the advantage comes from pretraining data versus the head design; the paper uses a grid search of learning rates but does not ablate head choices on the baselines.
  • The 2D-slice design leaves volumetric context on the table; the scaling curves suggest a native 3D extension of the same recipe, with more data or longer training, is the most direct path to further gains.
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 introduces Curia, a family of Vision Transformer foundation models (ViT-B and ViT-L) pre-trained with DINOv2 self-supervision on roughly 200 million 2D CT and MRI slices from 150,000 exams (130 TB) collected at a single private hospital between 2019 and 2022. The authors also release CuriaBench, a 19-task external benchmark covering organ recognition, registration, segmentation, classification, regression, and survival prediction across CT and MRI. Evaluation uses frozen backbones with lightweight heads and compares against MedImageInsight, BiomedCLIP, and, in places, DINOv2 and a cancer imaging foundation model. Claims include strong few-shot and cross-modality generalization, competitive or superior performance relative to resident radiologists across 14 tasks, and clinically useful emergent properties. The base model weights are released on Hugging Face.

Significance. If the central claims hold, this is a valuable contribution: it demonstrates that self-supervised pre-training on a large, routine clinical imaging corpus can yield a transferable radiology backbone, and CuriaBench is a potentially useful public benchmark. The statistical protocol is a strength: 5 training runs, 1000-sample bootstrapping, paired tests, and external public datasets. Release of model weights improves reproducibility. However, the headline claim of meeting or surpassing resident radiologists is not currently auditable, several main-text claims are contradicted by the paper's own supplementary tables, and a few data-reporting errors undermine confidence in the statistical presentation. These issues are fixable but require substantive revision.

major comments (5)
  1. [§4.4, Fig. 1e] The headline claim that Curia 'meets or surpasses the performance of radiologists' is not auditable. The text reports a mean over four final-year residents across 14 tasks but omits per-task test-subset sizes, per-reader scores, the metric used when averaging heterogeneous metrics (AUC, balanced accuracy, c-index, r²), and whether the paired bootstrap test compares Curia to each resident or to the resident average. Please provide a full per-task radiologist evaluation table and the exact protocol; without this, the abstract claim is unsupported.
  2. [§2.1, Table D6] The claim that Curia 'outperformed or matched the performance of other models across all image registration tasks' and 'outperformed others on all organ-specific metrics' is contradicted by Table D6. For XCAT CT→CT, BiomedCLIP has mean DSC 81.74 vs Curia-B 81.30 and Curia-L 80.12. For XCAT CT→MR, DINOv2 Large has higher liver DSC (86.22 vs 86.12 for Curia-B) and higher spleen DSC (71.90 vs 70.09 for Curia-B). Similarly, the Introduction's 'consistently and significantly outperforms existing foundation models' is contradicted by §2.4, where MedImageInsight significantly outperforms Curia-L on Abdominal Trauma (93.14 vs 87.10, P<0.001) and scores higher on Myocardial Infarction (94.08 vs 89.16). These claims must be qualified.
  3. [§2.5, Table E23] The text reports Curia-L AUROC 87.74 for the Alzheimer's disease task, but Table E23 reports 84.90 (95% CI 74.51–93.78). This is a direct numerical inconsistency. It must be resolved, since it affects the reported competitiveness in the neurodegenerative benchmark. The same table also lists Curia-B as 87.83, matching the text for Curia-B only.
  4. [§4.5.2] The Kidney Cancer Survival benchmark uses tumor volumes segmented by the authors' own segmentation FM [41], then 'the FM' is used with a Cox layer to predict time-to-event. Please specify whether the survival features are extracted from Curia or from the same segmentation FM, and describe the steps taken to prevent information leakage between mask generation and risk prediction. With n=183 patients, the c-index comparisons are also underpowered; the validation-set threshold selection should be described in more detail.
  5. [Appendix E, Tables E17 and E20] Two supplementary tables report inverted confidence intervals: Subarticular Stenosis Curia-B is 87.81 with lower 95% CI 89.09 and upper 95% CI 86.57 (Table E17), and Abdominal Trauma Curia-B is 82.63 with lower 95% CI 83.98 and upper 95% CI 81.31 (Table E20). These errors call into question the reliability of the other CI tables. Please correct them and re-check all bootstrap CI computations.
minor comments (6)
  1. [§2.1, Table E10] The r² score is reported as '75.54' and '69.41' etc. Since r² is normally bounded by 1, these values should be expressed as percentages (e.g., 75.54%) or as 0.7554 to avoid confusion.
  2. [§2.3] The heading 'Curia achieves leading performance in musculoskeletal disease assessment' overstates the results: for foraminal narrowing and spinal canal stenosis, Curia-L is statistically equivalent to MedImageInsight (P=0.87 and P=0.243, respectively). Consider rewording to 'comparable or better'.
  3. [Fig. 1e] The radiologist comparison is only described textually; the figure itself should display the per-task values with confidence intervals, or a dedicated table should be added, so that the 'average of four residents' result can be inspected.
  4. [§4.4] Please specify the level of the residents (e.g., final-year, but in which country's training system), whether they were supervised or board-certified, and how the 14 radiologist tasks were selected from the 19 benchmark tasks.
  5. [Table E11] The comparison with Harvard OncoFM is not apples-to-apples because OncoFM-finetuned updates the encoder, whereas Curia uses a frozen backbone. This is noted in the table caption, but it should also be prominently stated in the main text.
  6. [Fig. 5a caption] The caption mentions 'Harvard-RT', while the text and references refer to 'Harvard Onco-FM'. Please unify the nomenclature.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all benchmark tasks use external public datasets and the pre-training is independent of the benchmark labels.

full rationale

The paper's central claim is that self-supervised pre-training on 200M unlabeled routine CT/MRI images yields a transferable radiology backbone. This claim is evaluated on a 19-task benchmark built from external public datasets (TotalSegmentator, IXI, Learn2Reg, AMOS, LUNA16, KiTS23, DeepLesion, TCIA/TCGA, RSNA challenges, COVIDx, OASIS, EMIDEC, ATLAS). No task metric is computed from the pre-training data, and no prediction reduces by construction to a fitted parameter. The only self-citations are (i) RadSAM [21] for the prompted-segmentation protocol and as a baseline, and (ii) the segmentation FM [41] used to semi-automatically delineate kidney tumors before manual correction by two residents; both are ancillary and the final labels are human-corrected with external survival outcomes. The radiologist comparison in Fig. 1e, while under-audited, is an external human benchmark and does not enter the model's training or head selection. The statistical procedure (bootstrap, paired tests) is standard. Therefore no circularity step can be exhibited.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new particles, mediators, forces, or theoretical entities. Its free parameters are standard ML hyperparameters and one survival threshold. The key assumptions are domain-level: SSL transferability, label correctness, single-center representativeness, and no data leakage.

free parameters (3)
  • DINOv2 pre-training hyperparameters (batch size, iterations, warmup, weight decay, crop scales) = Curia-B: batch 512, 475k iterations, warmup 25k, weight decay 0.04 to 0.2; Curia-L: batch 256, same iterations and warmu
    Chosen by hand through smaller-dataset tuning; not fitted to the benchmark. These affect the learned features directly.
  • Downstream head configuration (10 learning rates, head type per task) = Not reported per task; selected on held-out validation set.
    Grid search over learning rate and head architecture (CLS token, patch pooling, attention pooling, Table B2). Standard model selection, but part of the reported performance.
  • Survival stratification threshold for Kaplan-Meier groups = Not reported.
    Section 4.2.1: "we chose a threshold that maximized the log-rank test statistic on a held-out validation set." Used for the KM curves in Fig. 5d; the c-index itself does not depend on this threshold.
assumptions (6)
  • domain assumption DINOv2 self-supervised pre-training on unlabeled 2D slices yields transferable anatomical features.
    The entire approach assumes that the SSL objective on 200M slices produces features useful for classification, regression, segmentation, and registration. No proof beyond the empirical results.
  • domain assumption Z-score normalization without modality- or task-specific windowing preserves diagnostic information in CT and MRI.
    Section 4.1 states all images were z-score normalized and no windowing was applied for Curia. If intensity standardization discards diagnostic signal, downstream results would be impaired.
  • domain assumption Ground truth labels in the public benchmark datasets (RSNA, LUNA16, KITS23, EMIDEC, ATLAS, OASIS, COVIDx, etc.) are correct.
    The benchmark scores are computed against these labels without independent pathology verification for most tasks.
  • standard math The Cox proportional hazards model is appropriate for the kidney cancer survival data.
    Section 4.2.1 uses torchsurv's Cox partial likelihood. Proportional hazards is a standard statistical assumption not tested in the paper.
  • domain assumption The single-center pre-training corpus is sufficiently representative of the broader radiology population.
    Acknowledged limitation in Section 3: institutional biases, vendor-specific protocols, and local patient population could affect generalizability.
  • domain assumption There is no patient overlap between the private pre-training hospital data and the public benchmark test sets.
    Pre-training data are anonymized and private, so overlap cannot be checked. If a benchmark patient were also in the pre-training corpus, results would be inflated. Not addressed in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Curia: A Multi-Modal Foundation Model for Radiology." pith.science (2026). https://pith.science/paper/2PP5LQZI

@misc{pith2026250906830,
  author       = {Pith},
  title        = {Pith review of: Curia: A Multi-Modal Foundation Model for Radiology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PP5LQZI}},
  note         = {Machine review of arXiv:2509.06830}
}
read the original abstract

AI-assisted radiological interpretation is based on predominantly narrow, single-task models. This approach is impractical for covering the vast spectrum of imaging modalities, diseases, and radiological findings. Foundation models (FMs) hold the promise of broad generalization across modalities and in low-data settings. However, this potential has remained largely unrealized in radiology. We introduce Curia, a foundation model trained on the entire cross-sectional imaging output of a major hospital over several years, which to our knowledge is the largest such corpus of real-world data-encompassing 150,000 exams (130 TB). On a newly curated 19-task external validation benchmark, Curia accurately identifies organs, detects conditions like brain hemorrhages and myocardial infarctions, and predicts outcomes in tumor staging. Curia meets or surpasses the performance of radiologists and recent foundation models, and exhibits clinically significant emergent properties in cross-modality, and low-data regimes. To accelerate progress, we release our base model's weights at https://huggingface.co/raidium/curia.

Discussion (0). Sign in to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Lesion Detection in CT with Frozen Self-Distilled Features: SALT, a Spatially Adaptive Label-Guided Temperature

    cs.CV 2026-08 conditional novelty 7.0 of 10

    Spatially adaptive teacher temperature and patch-loss weighting during self-distillation pretraining improve frozen-feature lesion separability and small-lesion detection in CT.

  2. Semantically Calibrated Evidence Composition for CT Vision-Language Learning

    cs.CV 2026-07 conditional novelty 6.0 of 10

    SCOPE composes organ-level CT evidence under whole-volume context and calibrates it with diagnostic-summary supervision, reaching 85.0/72.2 macro AUC on CT-RATE and RadChestCT.

  3. Learning How Much, Not Just What: Cross-Patient Burden Order for CT Vision-Language Pretraining

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Cross-patient report-based pair mining plus burden-direction alignment improves CT vision-language pretraining, reaching 85.6 AUROC on CT-RATE zero-shot diagnosis.

  4. Jolia: Concept-Level Vision-Language Alignment for 3D CT Contrastive Learning

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    ConQuer augments global CLIP alignment with independent per-concept contrastive losses on anatomical regions extracted from reports, producing Jolia which outperforms CLIP baselines on classification, report generatio...

  5. Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    FlexiCT applies agglomerative pretraining across 2D, 3D, and vision-language stages on a large public CT collection to produce representations that match or exceed specialized models on segmentation, classification, r...

  6. Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge

    cs.CV 2026-04 conditional novelty 6.0 of 10

    Self-supervised pretraining on 60K clinical-style brain MRIs improves out-of-domain generalization on classification, segmentation, and regression tasks, with hybrid objectives and small models showing strong results.

  7. Foundation Models vs. Radiomics for Lung Computed Tomography: A Benchmark of Feature Extractors, Classification Heads, and Segmentation Choices

    cs.CV 2026-07 conditional novelty 5.0 of 10

    Benchmark finds segmentation dominates volume and stage tasks while classifier choice dominates survival, histology, and age prediction, recommending Curia with tumor segmentation and CatBoost as a safe default.

  8. Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining

    cs.CV 2026-05 unverdicted novelty 5.0 of 10

    FlexiCT provides CT foundation models via agglomerative pretraining on 266227 volumes from 56 datasets that match or exceed task-specific models on five task families while organizing embeddings along tumor-stage gradients.

  9. Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge

    cs.CV 2026-04 accept novelty 5.0 of 10

    Self-supervised pretraining on large unlabeled clinical brain MRI data improves generalization to out-of-domain clinical tasks over supervised in-domain training, with task-specific optimal objectives and limited bene...

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages · cited by 7 Pith papers

  1. [41]

    URL https://doi.org/10

    Machado, L.et al.A promptable ct foundation model for solid tumor evaluation.npj Precision Oncology9, 121 (2025). URL https://doi.org/10. 1038/s41698-025-00903-y

    work page 2025

  2. [1]

    & Nawaz, D

    Baloch, M., Ali, K. & Nawaz, D. Illuminat- ing insights: The role of radiology in diagnosing medical conditions (2023)

    work page 2023

  3. [2]

    URL http://arxiv.org/abs/1702.05747

    Litjens, G.et al.A survey on deep learning in medical image analysis.CoRRabs/1702.05747 (2017). URL http://arxiv.org/abs/1702.05747

    arXiv 2017

  4. [3]

    URL https://www.mdpi.com/2075-4418/ 15/9/1146

    Nair, A.et al.Enhancing radiologist productivity with artificial intelligence in magnetic resonance imaging (mri): A narrative review.Diagnostics15 (2025). URL https://www.mdpi.com/2075-4418/ 15/9/1146

    work page 2025

  5. [4]

    Prevedello, L. M.et al.Challenges related to artificial intelligence research in medical imaging and the importance of image analysis competi- tions.Radiology: Artificial Intelligence1, e180031 (2019)

    work page 2019

  6. [5]

    & Yang, Q

    Bian, Y., Li, J., Ye, C., Jia, X. & Yang, Q. Arti- ficial intelligence in medical imaging: From task- specific models to large-scale foundation models. Chinese Medical Journal138, 651–663 (2025)

    work page 2025

  7. [6]

    Oquab, M.et al.Dinov2: Learning robust visual features without supervision (2023)

    work page 2023

  8. [7]

    He, K.et al.Masked autoencoders are scalable vision learners (2022)

    work page 2022

Show all 47 references
  1. [8]

    Paschali, M.et al.Foundation models in radiol- ogy: What, how, why, and why not.Radiology 314, e240597 (2025)

    2025

  2. [9]

    Lesaunier, A.et al.Artificial intelligence in inter- ventional radiology: Current concepts and future trends.Diagnostic and Interventional Imaging 106, 5–10 (2025)

    2025

  3. [10]

    Zhang, S.et al.BioMedCLIP: A Multimodal Biomedical Foundation Model Trained from Fif- teen Million Image–Text Pairs.NEJM AI2, AIoa2400640 (2025)

    2025

  4. [11]

    URL https://www.nature.com/ articles/s41592-024-02499-w

    Zhao, T.et al.A foundation model for joint seg- mentation, detection and recognition of biomedi- cal objects across nine modalities.Nature Meth- ods1–11 (2024). URL https://www.nature.com/ articles/s41592-024-02499-w. Publisher: Nature Publishing Group

    2024

  5. [13]

    A benchmarking crisis in biomed- ical machine learning.Nature Medicine31, 1060–1060 (2025)

    Mahmood, F. A benchmarking crisis in biomed- ical machine learning.Nature Medicine31, 1060–1060 (2025)

    2025

  6. [14]

    URL https://arxiv.org/abs/2010.11929

    Dosovitskiy, A.et al.An image is worth 16x16 words: Transformers for image recognition at scale.CoRRabs/2010.11929(2020). URL https://arxiv.org/abs/2010.11929

    2010 arXiv

  7. [15]

    Codella, N. C. F.et al.MedImageInsight: An Open-Source Embedding Model for General Domain Medical Imaging (2024). URL http: //arxiv.org/abs/2410.06542. ArXiv:2410.06542 [eess]

    2024 arXiv

  8. [16]

    Ding, M.et al.Davit: Dual attention vision transformers (2022)

    2022

  9. [17]

    URL https://ai.nejm

    Zhang, S.et al.BioMedCLIP: A Multimodal Biomedical Foundation Model Trained from Fif- teen Million Image–Text Pairs.NEJM AI 2, AIoa2400640 (2025). URL https://ai.nejm. org/doi/10.1056/AIoa2400640. Publisher: Mas- sachusetts Medical Society

    2025 doi

  10. [18]

    Hering, A.et al.Learn2reg: comprehensive multi- task medical image registration challenge, dataset and evaluation in the era of deep learning.IEEE Transactions on Medical Imaging42, 697–712 (2022)

    2022

  11. [19]

    URL https://arxiv.org/abs/2304.02643

    Kirillov, A.et al.Segment anything (2023). URL https://arxiv.org/abs/2304.02643. 2304.02643

    2023 arXiv

  12. [20]

    Nature Communications15(2024)

    Ma, J.et al.Segment anything in medical images. Nature Communications15(2024). URL http: //dx.doi.org/10.1038/s41467-024-44824-z

    2024 doi

  13. [21]

    Khlaut, J.et al.Radsam: Segmenting 3d radiolog- ical images with a 2d promptable model (2025)

    2025

  14. [22]

    URL https://www.nature.com/ articles/s42256-024-00807-9

    Pai, S.et al.Foundation model for cancer imag- ing biomarkers.Nature Machine Intelligence6, 354–367 (2024). URL https://www.nature.com/ articles/s42256-024-00807-9. Publisher: Nature Publishing Group

    2024

  15. [23]

    Swets, J.Evaluation of diagnostic systems(Else- vier, 2012)

    2012

  16. [24]

    & Adams, N

    Hogan, J. & Adams, N. M. On averaging roc curves.Transactions on Machine Learning Research(2023)

    2023

  17. [25]

    Pai, S.et al.Foundation model for cancer imaging biomarkers.Nature machine intelligence6, 354– 367 (2024)

    2024

  18. [26]

    Cox, D. R. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological)34, 187–202 (1972)

    1972

  19. [27]

    Monod, M.et al.torchsurv: a lightweight pack- age for deep survival analysis.arXiv preprint arXiv:2404.10761(2024)

    2024 arXiv

  20. [28]

    Clark, K.et al.The cancer imaging archive (tcia): maintaining and operating a public information repository.Journal of digital imaging26, 1045– 1057 (2013)

    2013

  21. [29]

    Marcus, D. S.et al.Open access series of imag- ing studies (oasis): cross-sectional mri data in young, middle aged, nondemented, and demented older adults.Journal of cognitive neuroscience19, 19 1498–1507 (2007)

    2007

  22. [30]

    & Bojanowski, P

    Moutakanni, T., Oquab, M., Szafraniec, M., Vakalopoulou, M. & Bojanowski, P. You don’t need domain-specific data augmentations when scaling self-supervised learning.Advances in Neu- ral Information Processing Systems37, 116106– 116125 (2024)

    2024

  23. [31]

    & Yan, P

    Song, X., Xu, X. & Yan, P. Dino-reg: General pur- pose image encoder for training-free multi-modal deformable medical image registration (2024)

    2024

  24. [32]

    URL http://dx.doi.org/10.1148/ryai.230024

    Wasserthal, J.et al.Totalsegmentator: Robust segmentation of 104 anatomic structures in ct images.Radiology: Artificial Intelligence5(2023). URL http://dx.doi.org/10.1148/ryai.230024

    2023 doi

  25. [33]

    A.et al.Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images.arXiv preprint arXiv:2405.19492(2024)

    D’Antonoli, T. A.et al.Totalsegmentator mri: Sequence-independent segmentation of 59 anatomical structures in mr images.arXiv preprint arXiv:2405.19492(2024)

    2024 arXiv

  26. [34]

    URL https://brain-development.org/ ixi-dataset

    Ixi dataset. URL https://brain-development.org/ ixi-dataset

  27. [35]

    Bauer, D. F.et al.Generation of annotated multimodal ground truth datasets for abdominal medical image registration.International jour- nal of computer assisted radiology and surgery16, 1277–1285 (2021)

    2021

  28. [36]

    Ji, Y.et al.Amos: A large-scale abdominal multi- organ benchmark for versatile medical image segmentation.Advances in neural information processing systems35, 36722–36732 (2022)

    2022

  29. [37]

    Setio, A. A. A.et al.Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge.Medical image analysis42, 1–13 (2017)

    2017

  30. [38]

    Heller, N.et al.The kits21 challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase ct (2023). 2307. 01984

    2023

  31. [39]

    & Summers, R

    Yan, K., Wang, X., Lu, L. & Summers, R. M. Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning.Journal of medical imaging5, 036501–036501 (2018)

    2018

  32. [40]

    Comprehensive genomic characterization defines human glioblastoma genes and core pathways

    The Cancer Genome Atlas Research Network. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature455, 1061–1068 (2008). URL https://doi. org/10.1038/nature07385

    2008 doi

  33. [42]

    https://kaggle.com/competitions/ rsna-2024-lumbar-spine-degenerative-classification (2024)

    Richards, T.et al.Rsna 2024 lumbar spine degenerative classifica- tion. https://kaggle.com/competitions/ rsna-2024-lumbar-spine-degenerative-classification (2024). Kaggle

    2024

  34. [43]

    & Uenal, G

    ˇStajduhar, I., Mamula, M., Mileti´ c, D. & Uenal, G. Semi-automated detection of anterior cruciate ligament injury from mri.Computer methods and programs in biomedicine140, 151–164 (2017)

    2017

  35. [44]

    Lalande, A.et al.Emidec: A database usable for the automatic evaluation of myocardial infarction from delayed-enhancement cardiac mri.Data5 (2020)

    2020

  36. [45]

    https://kaggle.com/competitions/ rsna-2023-abdominal-trauma-detection (2023)

    Colak, E.et al.Rsna 2023 abdominal trauma detection. https://kaggle.com/competitions/ rsna-2023-abdominal-trauma-detection (2023). Kaggle

    2023

  37. [46]

    E.et al.Construction of a machine learning dataset through collaboration: the rsna 2019 brain ct hemorrhage challenge.Radiology: Artificial Intelligence2, e190211 (2020)

    Flanders, A. E.et al.Construction of a machine learning dataset through collaboration: the rsna 2019 brain ct hemorrhage challenge.Radiology: Artificial Intelligence2, e190211 (2020)

    2019

  38. [47]

    URL https://fcon 1000

    Liew, S.-L.et al.Anatomical tracings of lesions after stroke (atlas) r2.0. URL https://fcon 1000. projects.nitrc.org/indi/retro/atlas.html

  39. [48]

    & Wong, A

    Gunraj, H., Sabri, A., Koff, D. & Wong, A. Covid-net ct-2: Enhanced deep neural net- works for detection of covid-19 from chest ct images through bigger, more diverse learning. Frontiers in Medicine8, 729287 (2022). URL https://www.frontiersin.org/articles/10.3389/ fmed.2021.7...

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.