Pith. sign in

REVIEW 4 major objections 7 minor 50 references

Inserting synthetic coronary plaques into unlabeled CT scans teaches a model to read real disease better than standard pretraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 18:33 UTC pith:5YJ7IQSC

load-bearing objection Solid multi-center empirical win for pathology-synthesis pretraining on CCTA; the engine is under-ablated but the external gains still make the paper worth engaging. the 4 major comments →

arxiv 2603.24847 v2 pith:5YJ7IQSC submitted 2026-03-25 cs.CV

CORA: Generalizable coronary artery disease assessment and risk stratification from coronary CT angiography using pathology-centric representation learning

Jinkui Hao , Gorkem Durak , Halil Ertugrul Aktas , Ulas Bagci , Bradley D. Allen , Nilay S. Shah , Bo Zhou This is my paper
classification cs.CV
keywords coronary CT angiographyself-supervised learningpathology synthesisplaque characterizationstenosis detectioncoronary segmentationMACE predictionfoundation model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Coronary CT angiography can show the plaques that cause heart attacks, but the plaques occupy a tiny fraction of each scan and expert labels are scarce. Standard self-supervised pretraining spends its capacity reconstructing the dominant background anatomy, so it under-learns the sparse disease signal. CORA flips the objective: an anatomy-guided engine inserts realistic-looking calcified and non-calcified synthetic lesions into unlabeled volumes, turning pretraining into an abnormality-detection task that forces the network to notice plaque-like features. Pretrained on more than ten thousand unlabeled scans and fine-tuned on multi-center data from nine hospitals, the same encoder then outperforms strong 3D baselines on plaque typing, stenosis detection, vessel segmentation, and short-term major adverse cardiac event risk, with the largest gains on external sites. Coupling imaging features with clinical text further improves near-term risk stratification. The practical claim is that deliberately teaching the model to hunt for synthetic pathology yields transferable, annotation-efficient representations for comprehensive coronary disease assessment.

Core claim

Pathology-centric, synthesis-driven self-supervised pretraining on large unlabeled CCTA produces representations that consistently outperform strong 3D self-supervised baselines and training from scratch across plaque characterization, stenosis detection, coronary artery segmentation, and multimodal 30-day MACE prediction, with the largest gains under multi-center distributional shift.

What carries the argument

An anatomy-guided lesion synthesis engine that inserts diverse synthetic calcified and non-calcified plaques (modeled as 1–3 overlapping Gaussian blobs with controlled HU ranges) into unlabeled CCTA volumes, reframing pretraining as a sparse abnormality-detection task and thereby biasing the encoder toward clinically relevant vascular pathology rather than background anatomy.

Load-bearing premise

That synthetic plaques made from a few overlapping Gaussian blobs with fixed density ranges are diverse enough to teach the model the real disease features that matter clinically.

What would settle it

If a controlled ablation that replaces the synthetic-lesion engine with ordinary masked reconstruction (or removes synthesis entirely) closes the multi-center performance gap on plaque typing, stenosis F1, and external MACE AUC, the pathology-centric claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single encoder can support plaque typing, stenosis localization, vessel segmentation, and short-term risk scoring without multi-stage vessel reformatting pipelines.
  • Unlabeled institutional CCTA archives become usable for pretraining without diagnostic labels or plaque curation.
  • Pathology-centric pretraining yields larger external-site gains than generic anatomical reconstruction, suggesting better robustness to scanner and protocol differences.
  • Multimodal fusion of the imaging encoder with clinical text improves 30-day MACE prediction beyond imaging alone, especially under population shift.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same synthesis-as-abnormality-detection idea may transfer to other sparse-pathology 3D tasks (e.g., small-lesion CT or MRI) where background anatomy dominates.
  • Because pretraining needs only vessel masks, not plaque labels, the pipeline can scale to any hospital archive that can run an off-the-shelf coronary segmenter.
  • Zero-shot abnormality maps already light up real plaques, so the pretrained encoder may serve as a sensitive candidate generator that later fine-tuning only has to prune.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. CORA is a 3D Residual U-Net foundation model for CCTA that replaces generic anatomical SSL (masked reconstruction / contrastive learning) with pathology-centric, synthesis-driven pretraining: an anatomy-guided engine inserts synthetic calcified (800–1500 HU) and non-calcified (30–90 HU) lesions—formed as 1–3 overlapping Gaussian blobs—into artery-centric patches from unlabeled volumes, with multi-window inputs and a Tversky+Focal detection objective. Pretrained on a large Northwestern multi-hospital CCTA archive (~12.8k volumes in the main text), the encoder is fine-tuned for volume-level multi-label plaque characterization, lesion-level stenosis detection, coronary artery segmentation (ImageCAS), and multimodal 30-day MACE prediction (imaging + frozen Qwen-7B clinical text). Across internal and external (eight-hospital) cohorts, CORA outperforms from-scratch training and three 3D SSL baselines (MAE, VolumeFusion, VoCo), with the largest margins under external distributional shift; Grad-CAM and zero-shot abnormality maps are offered as supporting evidence of pathology-focused attention.

Significance. If the pathology-centric pretraining advantage is real and isolable, this is a meaningful contribution to annotation-efficient cardiovascular imaging and, more broadly, to SSL for spatially sparse medical pathology. Strengths that should be credited include: multi-task clinical coverage (plaque, stenosis, segmentation, near-term MACE); a true multi-center external holdout with bootstrap CIs and significance testing; data-efficiency curves on ImageCAS; and an honest Discussion of synthetic-lesion limitations, retrospective MACE noise, and asymptomatic-population scope. The framing—that morphological/attenuation diversity can create a useful over-sensitive prior without full biological fidelity—is a clear, testable design thesis. Those claims would be substantially more persuasive with controlled ablations and a fully specified, matched pretraining protocol for the SSL baselines.

major comments (4)
  1. Central claim isolation (Methods “Data engine”; Discussion on synthetic fidelity; Figs. S2–S3): The paper’s load-bearing thesis is that synthesis-driven abnormality detection—not artery-centric sampling, multi-windowing, residual U-Net capacity, or scale alone—explains the gains over MAE/VolumeFusion/VoCo. No ablation removes or degrades the synthetic engine (e.g., random non-vascular intensity blobs; multi-window MAE on the same corpus; artery-centric sampling without lesion insertion; varying blob count/sigma/HU ranges; more complex morphologies). Zero-shot maps show high sensitivity with many false positives, which is consistent with over-sensitivity but does not attribute external AUROC/F1 gains to the specific synthetic distribution. Without these controls, the “pathology-centric” framing remains under-supported relative to the strength of the claim.
  2. Fairness of SSL baseline pretraining (Results opening; Methods “CORA implementation and pretraining protocol”): It is not stated whether MAE, VolumeFusion, and VoCo were re-pretrained from scratch on the same ~12.8k CCTA volumes under their native objectives (with matched compute, multi-window inputs, and artery-centric sampling where applicable) or used off-the-shelf weights from other domains/datasets. If the latter, external gains partly reflect domain match rather than objective design. Please specify the full pretraining recipe for each baseline (data, windows, sampling, epochs, architecture) so that the comparison isolates the synthesis objective.
  3. Stenosis detection generalizability (Results “Pathology-aware pretraining…”; Methods stenosis subsection): Unlike plaque and MACE, stenosis is evaluated on 348 volumes / 556 lesions without a stated multi-center external split, using a lenient lesion-level criterion (>10-voxel overlap). Given that the paper’s strongest narrative is robustness under distributional shift, either provide an external stenosis evaluation or clearly scope the stenosis claim as single-cohort and secondary. Also report precision–recall stratified by calcified vs non-calcified and by CAD-RADS grade, not only aggregate F1.
  4. Dependence on external artery labels in “label-free” pretraining (Methods Data engine; Discussion on label independence): Pretraining requires coronary masks from an nnU-Net trained on ImageCAS, plus TotalSegmentator heart crops for stenosis fine-tuning. The Discussion correctly notes independence from diagnostic plaque/MACE labels, but the pipeline is not free of anatomical supervision. Please quantify sensitivity of downstream metrics to mask quality (e.g., noisy/eroded masks, or a no-mask random-crop control) and state this dependency explicitly in the abstract/intro so “annotation-efficient / unlabeled” is not overstated.
minor comments (7)
  1. Pretraining cohort size is inconsistent: abstract states 10,138 unlabeled volumes; Introduction/Results/Methods state 12,801. Reconcile and use one number throughout.
  2. Patch size inconsistency: Methods Data engine describes 64×64×64 artery-centric crops; Table S1, Computational resources, and Discussion refer to 96×96×96. Clarify the actual pretraining crop size.
  3. Typos/clarity: “real-the-fly synthesis”; “Theand fed into”; “decision-making decision-making”; “V olumeFusion” / “V oCo” spacing artifacts; abstract vs title wording drift (pathology-centric representation learning vs “Pathology Synthesis Driven Foundation Model”).
  4. Table 1 and MACE: report missingness rates for labs/clinical fields and how the Qwen text encoder handles missing values; this affects multimodal external gains.
  5. Fig. 2e / S5 Grad-CAM: state layer used, whether maps are class-specific for calcified vs non-calcified heads, and whether attention was quantified (e.g., vessel-mask overlap) beyond qualitative examples.
  6. Code availability is a placeholder (“[CORA]”). For a foundation-model claim, commit to releasing pretraining code for the synthesis engine and fine-tuning configs; weights if IRB allows.
  7. Resampling to 0.5×0.5×5 mm³ is aggressive in z; briefly discuss impact on distal vessel continuity and non-calcified plaque detectability.

Circularity Check

0 steps flagged

No load-bearing circularity: pretraining targets are synthetic lesions independent of real downstream labels; empirical gains are measured against external baselines on held-out multi-center data.

full rationale

CORA's central claim is empirical: a synthesis-driven SSL objective (inserting synthetic calcified/non-calcified lesions generated as 1–3 Gaussian blobs with fixed HU ranges into artery-centric patches, then training a Residual U-Net as abnormality detector) produces encoder representations that, after fine-tuning, outperform MAE, VolumeFusion, VoCo, and from-scratch baselines on plaque multi-label classification, lesion-level stenosis detection, coronary segmentation (ImageCAS), and multimodal 30-day MACE. The pretraining supervisory signal is entirely synthetic and label-free with respect to the real clinical annotations used for fine-tuning and evaluation; those annotations come from radiology reports or expert lesion marking on separate cohorts (Northwestern internal/external, ImageCAS). No equation or definition makes a reported AUROC/F1/Dice equal to a fitted input by construction. Multi-windowing, Tversky+Focal loss, and artery-centric sampling are design choices, not circular reductions. The only minor shared element is that artery masks for synthesis sampling come from an nnU-Net trained on ImageCAS, which is also the segmentation benchmark; this is a weak shared anatomical prior, not a self-definitional loop or fitted-parameter-as-prediction, and does not force the plaque/stenosis/MACE gains. No uniqueness theorems, self-citation chains, or renamed known results underwrite the claims. The paper is self-contained against external baselines and held-out data; residual concerns about synthetic diversity are validity/ablation issues, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim rests on engineering and domain choices rather than free physical constants: synthetic plaque generators, imbalance losses, multi-window HU cutoffs, artery-centric sampling from a separate segmentation model, and retrospective clinical labels. No new physical entity is postulated; the 'invented' piece is the pathology-aware synthesis engine used as a pretraining prior.

free parameters (6)
  • Tversky loss α, β = α=0.1, β=0.9
    Set to α=0.1, β=0.9 to heavily penalize false negatives during pretraining; chosen by design, not derived.
  • Focal loss γ = γ=4.0
    Focusing parameter γ=4.0 down-weights easy background; hand-chosen hyperparameter of the pretraining objective.
  • Synthetic plaque HU ranges = calcified 800–1500; soft 30–90
    Calcified 800–1500 HU and soft 30–90 HU define the synthetic target appearance and thus the pretraining signal.
  • Gaussian blob lesion morphology params = 1–3 blobs, sigma 0.7–2.0
    1–3 overlapping Gaussians with sigma 0.7–2.0 generate lesion shapes; ad hoc generative choices.
  • Multi-window HU windows = (-100,140), (50,400), (350,700), (500,2000) HU
    Four clinical windows (fat/soft/angio/calcium) define the multi-channel input representation used throughout.
  • Patch size / artery-centric sampling = patches ~64–96³; Poisson I0=1e5, L=200 mm, σe=2.0
    64³ or 96³ artery-centered crops and noise/augmentation settings shape what the encoder sees in pretraining.
axioms (5)
  • domain assumption Coronary plaques are extremely spatially sparse (<~1% of volume), so global reconstruction SSL under-represents pathology.
    Stated in Introduction as the motivation for synthesis-driven pretraining; underpins why the new objective should help.
  • ad hoc to paper Synthetic lesions need morphological/attenuation diversity more than full biological fidelity to create a useful pathology prior.
    Explicit Discussion claim justifying Gaussian-blob synthesis; load-bearing for interpreting zero-shot maps and transfer gains.
  • domain assumption An ImageCAS-trained nnU-Net provides sufficiently accurate coronary masks to anchor artery-centric synthesis on unlabeled institutional CCTA.
    Methods Data engine; errors in masks would misplace synthetic lesions and weaken the pathology bias.
  • domain assumption Radiology-report-derived plaque labels and radiologist-annotated stenoses are adequate ground truth for multi-label and lesion-level evaluation.
    Methods study population / stenosis dataset; inter-reader variability is known but treated as usable supervision.
  • domain assumption 30-day MACE labels from the integrated health system sufficiently capture near-term events for prognostic evaluation.
    Authors acknowledge possible missing external/out-of-hospital events; still used as primary prognostic endpoint.
invented entities (2)
  • Anatomy-guided pathology synthesis engine (CORA data engine) no independent evidence
    purpose: Insert controlled synthetic calcified/non-calcified plaques into unlabeled CCTA to turn pretraining into abnormality detection.
    Core methodological invention of the paper; independent evidence is only the downstream empirical gains and qualitative zero-shot maps, not an external biological validation of the synthesizer.
  • CORA multimodal MACE head (CCTA encoder + frozen Qwen-7B clinical text) no independent evidence
    purpose: Fuse imaging representations with demographic/clinical variables for 30-day MACE risk stratification.
    Engineering composition of existing components; evidence is internal/external AUCs only.

pith-pipeline@v1.1.0-grok45 · 19364 in / 3770 out tokens · 38618 ms · 2026-07-13T18:33:57.279340+00:00 · methodology

0 comments
read the original abstract

Coronary artery disease, a leading cause of cardiovascular mortality worldwide, can be assessed non-invasively by coronary computed tomography angiography (CCTA). Although deep learning has advanced automated CCTA analysis, clinical translation remains constrained by the scarcity of expert-annotated data and by the spatial sparsity of coronary pathology, which occupies only a small fraction of each scan. Widely used label-free pretraining strategies, such as masked image modeling and contrastive learning, optimize for global anatomical reconstruction and tend to under-represent these tiny localized pathological features. Here we present CORA, an annotation-efficient model for comprehensive coronary artery disease assessment. Rather than reconstructing background anatomy, CORA learns from volumetric CCTA through a synthesis-driven self-supervised strategy: an anatomy-guided engine inserts diverse synthetic calcified and non-calcified lesions into unlabeled scans, reframing pretraining as an abnormality-detection task that biases representation learning toward clinically relevant disease features. We pretrained CORA on 10,138 unlabeled CCTA volumes and evaluated it across datasets from nine independent hospitals. Across plaque characterization, stenosis detection, and coronary artery segmentation, CORA consistently outperformed strong self-supervised pretraining baselines, with the largest gains on external multi-center data, indicating robust generalization under distributional shift. Coupling the imaging encoder with structured clinical variables further enabled near-term major adverse cardiac event (MACE) risk stratification. Our results show that pathology-centric, synthesis-driven pretraining is an effective and scalable strategy for annotation-efficient coronary artery disease assessment from CCTA.

Figures

Figures reproduced from arXiv: 2603.24847 by Bo Zhou, Bradley D. Allen, Gorkem Durak, Halil Ertugrul Aktas, Jinkui Hao, Nilay S. Shah, Ulas Bagci.

Figure 1
Figure 1. Figure 1: Overview of CORA and synthesis-driven pretraining framework. (a) Synthesis-driven self-supervised pretraining of CORA on large-scale unlabeled CCTA volumes. An anatomy-guided lesion synthesis engine generates diverse calcified and non-calcified plaque patterns with controlled morphology and attenuation, which are inserted into CCTA volumes to create simulated abnormality. Multi-windowing inputs capture com… view at source ↗
Figure 2
Figure 2. Figure 2: Performance of plaque characterization across internal and external cohorts. Receiver operating characteristic (ROC) curves for the detection of calcified and non-calcified coronary plaques on the internal test cohort (a,b) and external multi-center cohort (c,d). CORA consistently outperformed models trained from scratch and existing 3D foundation models, with the largest performance gains observed on the … view at source ↗
Figure 3
Figure 3. Figure 3: CORA performance in coronary stenosis detection. (a) Representative axial CCTA images illustrating the target pathologies. Magnified insets detail a low-contrast non-calcified stenosis (left, blue) and a calcified stenosis (right, red). (b) Distribution of annotated stenoses in the entire fine-tuning dataset, stratified by CAD-RADS severity grade and plaque composition (calcified, red; non-calcified, blue)… view at source ↗
Figure 4
Figure 4. Figure 4: Coronary artery segmentation performance and data efficiency. (a) Segmentation performance on the ImageCAS test set using 100 labeled training samples, evaluated by Dice similarity coefficient, centerline Dice (clDice), and mean surface distance (MSD). CORA outperformed models trained from scratch and existing 3D foundation models across all metrics. (b) Data efficiency analysis across varying numbers of l… view at source ↗
Figure 5
Figure 5. Figure 5: CORA performance in MACE prediction. Receiver operating characteristic (ROC) curves for MACE risk stratification on the internal (a) and external (b) validation datasets. The vision-only CORA(Image) (light red) consistently outperforms the supervised baseline and state-of-the-art 3D foundation models (MAE, VolumeFusion, VoCo). Fusing clinical metadata with imaging representations, CORA(Multi-modal) (solid … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 2 linked inside Pith

  1. [1]

    Lindstrom, M.et al.Global burden of cardiovascular diseases and risks collaboration, 1990-2021.J. Am. Coll. Cardiol.80, 2372–2425 (2022)

    1990

  2. [2]

    Members, W. C.et al.2021 aha/acc/ase/chest/saem/scct/scmr guideline for the evaluation and diagnosis of chest pain: a report of the american college of cardiology/american heart association joint committee on clinical practice guidelines.J. Am. Coll. Cardiol.78, e187–e285 (2021)

    2021

  3. [3]

    Narula, J.et al.Prospective deep learning–based quantitative assessment of coronary plaque by computed tomography angiography compared with intravascular ultrasound: the revealplaque study.Eur. Hear. Journal- Cardiovascular Imaging25, 1287–1295 (2024)

    2024

  4. [4]

    A.et al.Diagnostic accuracy of non-invasive cardiac imaging modalities in patients with a history of coronary artery disease: a meta-analysis.Heart111, 4–10 (2025)

    Jukema, R. A.et al.Diagnostic accuracy of non-invasive cardiac imaging modalities in patients with a history of coronary artery disease: a meta-analysis.Heart111, 4–10 (2025)

    2025

  5. [5]

    Williams, M. C.et al.Low-attenuation noncalcified plaque on coronary computed tomography angiography predicts myocardial infarction: results from the multicenter scot-heart trial (scottish computed tomography of the heart).Circulation141, 1452–1462 (2020)

    2020

  6. [6]

    M.et al.Ai-quantitative ct coronary plaque features associate with a higher relative risk in women: Confirm2 registry.Circ

    Feuchtner, G. M.et al.Ai-quantitative ct coronary plaque features associate with a higher relative risk in women: Confirm2 registry.Circ. Cardiovasc. Imaging18, e018235 (2025)

    2025

  7. [7]

    Coronary ct angiography and 5-year risk of myocardial infarction.New Engl

    Investigators, S.-H. Coronary ct angiography and 5-year risk of myocardial infarction.New Engl. J. Medicine379, 924–933 (2018)

    2018

  8. [8]

    S.et al.Outcomes of anatomical versus functional testing for coronary artery disease.New Engl

    Douglas, P. S.et al.Outcomes of anatomical versus functional testing for coronary artery disease.New Engl. J. Medicine372, 1291–1300 (2015)

    2015

  9. [9]

    Heal.4, e256–e265 (2022)

    Lin, A.et al.Deep learning-enabled coronary ct angiography for plaque and stenosis quantification and cardiac risk prediction: an international multicentre study.The Lancet Digit. Heal.4, e256–e265 (2022)

    2022

  10. [10]

    & Abdulla, J

    Thomsen, C. & Abdulla, J. Characteristics of high-risk coronary plaques identified by computed tomographic angiography and associated prognosis: a systematic review and meta-analysis.Eur. Hear. Journal-Cardiovascular Imaging17, 120–129 (2016)

    2016

  11. [11]

    S.et al.Development and validation of a quantitative coronary ct angiography model for diagnosis of vessel-specific coronary ischemia.Cardiovasc

    Nurmohamed, N. S.et al.Development and validation of a quantitative coronary ct angiography model for diagnosis of vessel-specific coronary ischemia.Cardiovasc. Imaging17, 894–906 (2024)

    2024

  12. [12]

    Diagnostics15, 1822 (2025)

    Pinna, A.et al.Machine learning for coronary plaque characterization: A multimodal review of oct, ivus, and ccta. Diagnostics15, 1822 (2025)

    2025

  13. [13]

    Shrivastava, P.et al.A systematic review on deep learning-enabled coronary ct angiography for plaque and stenosis quantification and cardiac risk prediction.Eur. J. Radiol. Open14, 100652 (2025)

    2025

  14. [14]

    & Hoffmann, U

    Maurovich-Horvat, P., Ferencik, M., V oros, S., Merkely, B. & Hoffmann, U. Comprehensive plaque assessment by coronary ct angiography.Nat. Rev. Cardiol.11, 390–402 (2014)

    2014

  15. [15]

    & Wang, K

    Ma, X., Luo, G., Wang, W. & Wang, K. Transformer network for significant stenosis detection in ccta of coronary arteries. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 516–525 (Springer, 2021)

    2021

  16. [16]

    imaging13, 1409–1417 (2020)

    Lee, S.-E.et al.Differences in progression to obstructive lesions per high-risk plaque features and plaque volumes with ccta.Cardiovasc. imaging13, 1409–1417 (2020)

    2020

  17. [17]

    In International Conference on Medical Image Computing and Computer-Assisted Intervention, 593–601 (Springer, 2019)

    Denzinger, F.et al.Coronary artery plaque characterization from ccta scans using deep learning and radiomics. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 593–601 (Springer, 2019). 16/18

    2019

  18. [18]

    Lorenzatti, D.et al.Interaction between risk factors, coronary calcium, and ccta plaque characteristics in patients aged 18–45 years.Eur. Hear. Journal-Cardiovascular Imaging25, 1071–1082 (2024)

    2024

  19. [19]

    Dey, D.et al.Automated three-dimensional quantification of noncalcified coronary plaque from coronary ct angiography: comparison with intravascular us.Radiology257, 516–522 (2010)

    2010

  20. [20]

    Zreik, M.et al.A recurrent cnn for automatic detection and classification of coronary artery plaque and stenosis in coronary ct angiography.IEEE transactions on medical imaging38, 1588–1598 (2018)

    2018

  21. [21]

    Du, T.et al.Training and validation of a deep learning architecture for the automatic analysis of coronary angiography: Automatic recognition of coronary angiography.EuroIntervention17, 32 (2021)

    2021

  22. [22]

    A.et al.Artificial intelligence–based coronary plaque quantification using coronary ct angiography: Current insights and future directions.Radiol

    Jukema, R. A.et al.Artificial intelligence–based coronary plaque quantification using coronary ct angiography: Current insights and future directions.Radiol. Cardiothorac. Imaging7, e240568 (2025)

    2025

  23. [23]

    R.et al.Coronary artery stenosis and high-risk plaque assessed with an unsupervised fully automated deep learning technique.JACC: Adv.3, 100861 (2024)

    Ihdayhid, A. R.et al.Coronary artery stenosis and high-risk plaque assessed with an unsupervised fully automated deep learning technique.JACC: Adv.3, 100861 (2024)

    2024

  24. [24]

    Intell.46, 9052–9071 (2024)

    Gui, J.et al.A survey on self-supervised learning: Algorithms, applications, and future trends.IEEE Transactions on Pattern Analysis Mach. Intell.46, 9052–9071 (2024)

    2024

  25. [25]

    & Metaxas, D

    Zhang, S. & Metaxas, D. On the challenges and perspectives of foundation models for medical image analysis. Med. image analysis91, 102996 (2024)

    2024

  26. [26]

    Intell.47, 2245–2264 (2025)

    Awais, M.et al.Foundation models defining a new era in vision: a survey and outlook.IEEE Transactions on Pattern Analysis Mach. Intell.47, 2245–2264 (2025)

    2025

  27. [27]

    Zhou, Y .et al.A foundation model for generalizable disease detection from retinal images.Nature622, 156–163 (2023)

    2023

  28. [28]

    J.et al.Towards a general-purpose foundation model for computational pathology.Nat

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology.Nat. medicine30, 850–862 (2024)

    2024

  29. [29]

    Wang, X.et al.A pathology foundation model for cancer diagnosis and prognosis prediction.Nature634, 970–978 (2024)

    2024

  30. [30]

    & Yan, B

    Ma, C., Tan, W., He, R. & Yan, B. Pretraining a foundation model for generalizable fluorescence microscopy-based image restoration.Nat. Methods21, 1558–1567 (2024)

    2024

  31. [31]

    methods1–14 (2025)

    Tejada-Lapuerta, A.et al.Nicheformer: a foundation model for single-cell and spatial omics.Nat. methods1–14 (2025)

    2025

  32. [32]

    Medicine8, 678 (2025)

    Yao, J.et al.Eva-x: A foundation model for general chest x-ray analysis with self-supervised learning.npj Digit. Medicine8, 678 (2025)

    2025

  33. [33]

    Zhou, Z., Sodha, V ., Pang, J., Gotway, M. B. & Liang, J. Models genesis.Med. image analysis67, 101840 (2021)

    2021

  34. [34]

    InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 1970–1980 (2023)

    Chen, Z.et al.Masked image modeling advances 3d medical image analysis. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 1970–1980 (2023)

    1970

  35. [35]

    Imaging41, 2658–2669 (2022)

    Yan, K.et al.Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images.IEEE Transactions on Med. Imaging41, 2658–2669 (2022)

    2022

  36. [36]

    InProceedings of the Computer Vision and Pattern Recognition Conference, 5186–5196 (2025)

    Wald, T.et al.Revisiting mae pre-training for 3d medical image segmentation. InProceedings of the Computer Vision and Pattern Recognition Conference, 5186–5196 (2025)

    2025

  37. [37]

    Syst.(2025)

    Wang, H.et al.Sam-med3d: A vision foundation model for general-purpose segmentation on volumetric medical images.IEEE Transactions on Neural Networks Learn. Syst.(2025)

    2025

  38. [38]

    InProceedings of the IEEE/CVF International Conference on Computer Vision, 23839–23879 (2025)

    Wald, T.et al.An openmind for 3d medical vision self-supervised learning. InProceedings of the IEEE/CVF International Conference on Computer Vision, 23839–23879 (2025)

    2025

  39. [39]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16000–16009 (2022)

    He, K.et al.Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16000–16009 (2022)

    2022

  40. [40]

    Wang, G.et al.Mis-fm: 3d medical image segmentation using foundation models pretrained on a large-scale unannotated dataset.arXiv preprint arXiv:2306.16925(2023)

    Pith/arXiv arXiv 2023

  41. [41]

    & Chen, H

    Wu, L., Zhuang, J. & Chen, H. Large-scale 3d medical image pre-training with geometric context priors.IEEE Transactions on Pattern Analysis Mach. Intell.(2025). 17/18

    2025

  42. [42]

    R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization

    Selvaraju, R. R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, 618–626 (2017)

    2017

  43. [43]

    Zeng, A.et al.Imagecas: A large-scale dataset and benchmark for coronary artery segmentation based on computed tomography angiography images.Comput. Med. Imaging Graph.109, 102287 (2023)

    2023

  44. [44]

    Schlett, C. L.et al.Prognostic value of ct angiography for major adverse cardiac events in patients with acute chest pain from the emergency department: 2-year outcomes of the romicat trial.JACC: Cardiovasc. Imaging4, 481–491 (2011)

    2011

  45. [45]

    Imaging5, 990–999 (2012)

    Hou, Z.-h.et al.Prognostic value of coronary ct angiography and calcium score for major adverse cardiac events in outpatients.JACC: Cardiovasc. Imaging5, 990–999 (2012)

    2012

  46. [46]

    Imaging18, 308–319 (2025)

    Kolossváry, M.et al.Coronary plaque radiomic phenotypes predict fatal or nonfatal myocardial infarction: analysis of the scot-heart trial.Cardiovasc. Imaging18, 308–319 (2025). 47.Yang, A.et al.Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  47. [47]

    F., Kohl, S

    Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J. & Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nat. methods18, 203–211 (2021)

    2021

  48. [48]

    Salehi, S. S. M., Erdogmus, D. & Gholipour, A. Tversky loss function for image segmentation using 3d fully convolutional deep networks. InInternational workshop on machine learning in medical imaging, 379–387 (Springer, 2017)

    2017

  49. [49]

    & Dollár, P

    Lin, T.-Y ., Goyal, P., Girshick, R., He, K. & Dollár, P. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, 2980–2988 (2017)

    2017

  50. [50]

    Wasserthal, J.et al.Totalsegmentator: robust segmentation of 104 anatomic structures in ct images.Radiol. Artif. Intell.5, e230024 (2023). 18/18

    2023