Pith. sign in

REVIEW 3 major objections 5 minor 58 references

OrganLens claims a shared CT encoder can be conditioned on a selected organ identity to produce 11 organ-specific representations from one volume, with no external segmentation masks at inference.

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 ↗

OrganLens conditions a shared CT encoder on an organ identity and uses mask-supervised pooling to produce 11 organ-specific representations from the same volume, with organ-matched representations improving downstream detection and prognosis.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Serious, well-executed organ-specific CT representation paper with a real mechanism gap: mask quality on external cohorts is never measured, so the anatomy-weighted pooling claim rests on indirect ablations. the 3 major comments →

arxiv 2607.25164 v1 pith:ERYZJRXY submitted 2026-07-28 cs.CV cs.AI

OrganLens: Organ-Specific Representation Learning for CT Foundation Models

Zhixuan Ge , Anqi Li , Sadeer Al-Kindi , Hanwen Xu , Wei Qiu This is my paper
classification cs.CV cs.AI
keywords CT foundation modelsorgan-specific representation learningself-supervised learningmedical image analysisanatomy-weighted poolingorgan identity conditioningchest CTprognostic prediction
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.

The reading

The paper sets out to solve a mismatch: one CT scan contains many organs, but standard CT foundation models return a single volume-level representation that blends them together. OrganLens claims a single shared encoder can be asked, at inference, for the representation of a specific organ, and it will adapt its features accordingly — yielding 11 organ-specific representations from the same volume without needing masks. The paper shows that when a downstream task is organ-focused, the anatomically matched representation carries the strongest signal: heart-matched features raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953, and lung-matched features improve NLST lung-cancer mortality C-index from 0.683 to 0.780. A concatenated global representation keeps broad utility, including best image–text retrieval on two cohorts. If correct, this gives clinical researchers organ-indexed CT features without per-scan segmentation.

Core claim

The central claim is that one CT-pretrained vision transformer can be made organ-specific by adding a learned organ-identity offset to its CLS token, then trained with organ-guided crop sampling, self-distillation, and supervision to predict the selected organ's spatial mask. At inference the predicted mask is used only as soft weights: patch features are pooled within a slice, and slices are pooled in proportion to predicted organ area. The paper reports that this organ-conditioned pooling produces representations that outperform volume-level baselines for organ-specific abnormality detection and prognosis across CT-RATE, RAD-ChestCT, INSPECT, and NLST, while concatenating the 11 organ-spec

What carries the argument

The load-bearing mechanism is organ-identity conditioning plus anatomy-weighted pooling. A learned embedding for each of 11 organs is added (scaled by a trainable alpha initialized at zero) to the CLS token before the patch sequence enters the shared transformer, so the same backbone adapts per organ. During pretraining, a decoder supervised by BCE plus Dice on pseudo-labels produces a spatial mask; at inference, the sigmoid of this mask's logits provides soft weights for pooling patch features per slice and for weighting slices by predicted organ area. The mask is never supplied externally at inference.

Load-bearing premise

The load-bearing assumption is that the decoder's soft anatomical masks, trained only on CT-RATE pseudo-labels, remain spatially reliable enough on new scanners and contrast protocols to serve as pooling weights; the paper reports no mask-quality metrics on RAD-ChestCT, INSPECT, or NLST and itself notes that pseudo-label errors may propagate into the predicted mask.

What would settle it

On RAD-ChestCT, INSPECT, or NLST, hold out segmentations and compute Dice between predicted and true organ masks; then replace predicted masks at inference with oracle masks. If Dice is low, or if the AUROC/C-index gains over volume-level baselines disappear with oracle masks, the claimed advantage is coming from anatomy-weighted pooling rather than organ-conditioned features.

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

If this is right

  • For organ-focused tasks, use the anatomically matched representation; the paper reports it beats volume-level baselines on nearly all tested abnormalities and prognostic endpoints.
  • One model covers organ-specific and whole-volume needs: choose an organ identity for focused analysis, or concatenate all 11 for global tasks.
  • Inference requires only the CT volume and an organ label, so the approach can be dropped into existing CT-pretrained pipelines without a segmentation preprocessing step.
  • Because each slice is weighted by predicted organ area, small or sparsely sampled organs are less likely to be drowned out by large organs.
  • The 11 representations tie predictive signal to specific anatomy, supporting organ-level interpretation of what the model is responding to.

Where Pith is reading between the lines

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

  • If the predicted masks transfer reliably, the same recipe should work for abdominal and pelvic CT by extending the organ bank beyond thoracic organs; nothing in the method is inherently chest-specific.
  • Organ-conditioned features could be aligned with organ-specific report text, turning the retrieval anchor into a more precise clinical signal — the paper's global-retrieval result hints at this but does not build it.
  • A decisive extension would be to measure predicted-mask Dice on the transfer cohorts and test whether oracle masks recover the reported gains, separating the conditioning effect from the mask-pooling effect.
  • The longitudinal NLST results suggest organ-conditioned representations could support organ-level trajectory monitoring (e.g., a lung-risk score over screening visits), though the paper only reports repeated cross-sectional estimates.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper introduces OrganLens, a CT foundation-model adaptation that produces 11 organ-specific representations from a single shared ViT encoder. Organ identity is injected as an additive CLS-token offset; organ-guided cropping and self-distillation shape features; and an auxiliary anatomy decoder is trained with TotalSegmentator pseudo-labels from CT-RATE. At inference, the decoder predicts soft spatial masks that weight patch features within slices and slices within the volume (Eqs. 5–6), yielding organ-specific representations without external segmentation. The authors evaluate on CT-RATE, RAD-ChestCT, INSPECT, and NLST across abnormality detection, prognosis, longitudinal risk, and image–text retrieval, reporting that anatomically matched representations outperform volume-level baselines and that the global concatenated representation retains strong retrieval performance. The paper also includes component-level ablations, pooling ablations, and an explicit limitations section.

Significance. If the central claim holds, OrganLens addresses a real and under-served problem: many clinical questions target a specific organ inside a CT volume, while standard CT foundation models emit a single volume-level representation. The proposed design—conditioning a shared encoder on an organ identity and using a self-predicted mask for anatomy-weighted pooling—is scalable and does not require external segmentation at inference. The paper has notable strengths: evaluation across four independent cohorts and four task families; predefined anatomical matches rather than post hoc best-organ selection; transparent ablations showing that mask supervision and anatomy-weighted pooling contribute; and honest discussion of limitations. The main risk is that the inference-time spatial mask—the mechanism that makes the representations organ-specific—is never directly validated on the three non-CT-RATE cohorts. The comparison design also leaves open how much of the observed gain comes from organ conditioning versus simply adapting the backbone on CT-RATE. These concerns are addressable with additional experiments, but they affect attribution of the core mechanism.

major comments (3)
  1. [§3.4–3.5, §7 Limitations] The anatomy decoder is trained exclusively on CT-RATE TotalSegmentator pseudo-labels (Sec. 4.2, Eq. 3), yet at inference its soft masks are the pooling weights in Eqs. (5)–(6) for RAD-ChestCT, INSPECT, and NLST. No mask-quality metric (Dice, organ-area correlation, visual inspection rate) is reported on any of the three transfer cohorts. The Limitations sentence 'TotalSegmentator pseudo-label errors may propagate into the predicted spatial mask' addresses training-label noise, not cross-scanner mask reliability. Because the central claim is that organ-specific representations are produced by anatomy-weighted pooling—not merely by organ-identity conditioning—the current evidence does not rule out that on external acquisitions the masks are spatially poor and the method degrades to a conditioning-biased global average. The downstream ablations (Fig. 6) are suggestive but indirect. I reques
  2. [§5.5, Table S4; §4.3] The cleanest ablation ('w/o Organ-guided Crops and Anatomy-Mask Supervision') still includes organ-identity conditioning and three passes of adaptation on CT-RATE. It already improves over frozen CT-pretrained DINOv2 (0.798 vs 0.788 on CT-RATE; 0.711 vs 0.684 on RAD-ChestCT). Therefore the headline gains of full OrganLens over DINOv2 conflate (i) continued pretraining on CT-RATE, (ii) organ-identity conditioning, and (iii) anatomy-mask supervision. A control that adapts the same backbone on CT-RATE with the identical DINOv2 objective but no organ conditioning is needed to attribute the improvements to the organ-specific mechanism. Without this control, the abstract's statement that the heart representation 'raises CT-RATE cardiomegaly AUROC from 0.910 to 0.953' is not shown to be uniquely due to organ conditioning rather than to additional in-cohort adaptation.
  3. [§5.1–5.2, Supp. Tables S1–S2] The main results for OrganLens use the anatomically matched organ representation, which is a pre-specified but not automatically derived choice. The paper reports full matrices, which is good, but the summary numbers (macro-AUROC 0.856 on CT-RATE, 0.767 on RAD-ChestCT; C-indices 0.741 and 0.789) are computed after selecting the matched organ for each endpoint. Since the user of the model must know the target organ in advance, the practical utility is clear; however, the comparison to volume-level baselines would be fairer if it also reported the performance of an automatically selected representation (e.g., via a validation-set organ selector) or the mean performance across all 11 organs. The current framing can overstate the advantage of the method when the target organ is unknown.
minor comments (5)
  1. [Figure 2] The rank heatmaps in Figure 2a are dense and hard to read. Please add explicit best-organ labels or a companion table with the matched-organ rank for each abnormality.
  2. [§5.5 / Supp. Table S5] The pooling ablation is valuable, but some differences are within 0.01 AUROC (e.g., CT-RATE 'Uniform Patch Pooling' vs 'w/o Inter-slice Weighting': 0.832 vs 0.851). Please report confidence intervals or bootstrap significance tests for the key macro-AUROC comparisons.
  3. [§5.2, Supp. Table S2] The NLST endpoint labels 'LC inc.', 'All-cause LC death', and 'LC death' are ambiguous. Define each endpoint precisely, ideally with the exact event definitions used in the NLST data dictionary.
  4. [§4.2 / Supp. A.2] Please state explicitly whether the baseline representations (CT-pretrained DINOv2, GigaHeart, Merlin, SPECTRE) use the same 64-slice uniform sampling and 224×224 preprocessing as OrganLens. If not, differences in preprocessing could confound the comparison.
  5. [§5.4] The retrieval result for the 'global' representation is presented as a single concatenated vector. Since Supp. Table S3 shows that mean aggregation performs nearly as well, the text should avoid implying that concatenation specifically is essential; the key finding is that the 11-organ set preserves retrieval utility.

Circularity Check

0 steps flagged

No significant circularity: representation equations are explicit, downstream selection is a priori, and the GigaHeart self-citation is not load-bearing.

full rationale

OrganLens's derivation is self-contained in the relevant sense. The organ-specific representation z(V,o) is explicitly defined in Eqs. (5)-(6) as mask-weighted pooling of organ-conditioned patch features; the mask decoder is trained with the BCE+Dice objective in Eq. (3) against CT-RATE TotalSegmentator pseudo-labels and is then used at inference. No downstream result is fed back into the pretraining objective, and no reported 'prediction' is a fitted parameter renamed. The 'anatomically matched' evaluation selects representations by predefined organ-abnormality mappings (e.g., cardiomegaly->heart) before looking at AUROC/C-index; this is a protocol choice, not a performance-based selection that would make the matched gain an identity. The only self-citation (GigaHeart, [43]) provides the initial CT-pretrained DINOv2 checkpoint and a baseline; it does not supply a uniqueness theorem or forbid alternatives, and the central organ-specific mechanism is evaluated against external cohorts and ablations. The limitation about TotalSegmentator pseudo-label propagation concerns mask quality on transfer cohorts, which is an unverified empirical risk, not a circularity: even if the masks were unreliable, the conditioning mechanism would still be fitted, but no equation reduces the reported downstream gains to the fit by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical or biological entity is introduced. The 'organ token' is an additive offset to the CLS token, not a new sequence token or external entity; the predicted anatomy mask is a decoder output computed from input patches. The main external inputs the reader must accept are TotalSegmentator pseudo-labels, the GigaHeart-initialized DINOv2 backbone, and the hand-set hyperparameters listed above.

free parameters (4)
  • Organ identity embeddings e_o = 11 learned vectors, values not reported
    Central conditioning mechanism in Eq. (1); each organ's offset to the CLS token is learned during pretraining and is never reported or externally validated.
  • Organ-conditioning scale alpha = learned; initialized to 0
    Eq. (1) uses alpha e_o as an additive CLS offset; initialization preserves the pretrained model, but the final learned strength is not reported.
  • KoLeo loss weight lambda_KoLeo = 0.1
    Hand-set in Eq. (4); no sensitivity analysis is reported, and it directly shapes the learned representation.
  • Organ-guided crop acceptance thresholds = global: >=32 mask px, >=1% foreground, >=75% bbox; local: >=8 px, >=3% foreground
    Appendix A.3; these hand-set thresholds determine which views enter student-teacher distillation and therefore what the organ conditioner learns.
axioms (5)
  • domain assumption TotalSegmentator pseudo-labels on CT-RATE are an adequate proxy for organ ground-truth masks.
    The masks are used as y_o in Eq. (3) and for crop sampling in Section 3.3. The paper itself flags in Limitations that pseudo-label errors may propagate into predicted masks.
  • domain assumption 64 uniformly sampled 2D axial slices retain enough organ signal for all downstream tasks.
    Every representation is computed from 64 axial slices (Appendix A.2); the paper acknowledges that 2D encoding does not model volumetric continuity and may miss small findings.
  • ad hoc to paper A learned additive offset to the CLS token is sufficient to make a shared ViT produce organ-conditioned features.
    Eq. (1) defines conditioning as [c + alpha e_o]; no alternative conditioning scheme is compared, so the whole method depends on this design choice.
  • domain assumption Downstream labels are accurate enough to support the reported metrics.
    RAD-ChestCT labels were automatically extracted from reports, INSPECT and NLST use registry outcomes; the paper does not audit label quality.
  • domain assumption The CT-pretrained DINOv2 checkpoint released with GigaHeart is a suitable initialization.
    Student and teacher backbones are initialized from this external checkpoint (Section 4.3); if that checkpoint is weak or domain-mismatched, all results shift.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OrganLens: Organ-Specific Representation Learning for CT Foundation Models." pith.science (2026). https://pith.science/paper/ERYZJRXY

@misc{pith2026260725164,
  author       = {Pith},
  title        = {Pith review of: OrganLens: Organ-Specific Representation Learning for CT Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERYZJRXY}},
  note         = {Machine review of arXiv:2607.25164}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A CT examination captures multiple organs, but many biomedical questions concern abnormalities, prognosis, or longitudinal change in a specific organ. These questions require a separate representation for each organ within the same CT volume. Existing CT foundation models commonly produce a single volume-level representation, while recent anatomy-aware methods either encode pre-separated organ volumes or explicitly disentangle images into organ token groups. The former may remove clinically relevant surrounding context, while the latter does not condition a shared encoder on a selected organ before its features are formed. We introduce OrganLens for organ-specific representation learning through self-supervision. An organ identity conditions a shared CT encoder, while organ-specific distillation and anatomy-mask supervision shape features for anatomy-weighted pooling into organ-specific representations. At inference, the shared model produces 11 organ-specific representations without external segmentation masks. We evaluate OrganLens on CT-RATE, RAD-ChestCT, INSPECT, and NLST across diverse acquisitions and downstream evaluations. Relative to CT-pretrained DINOv2, heart representations raise CT-RATE cardiomegaly AUROC from 0.910 to 0.953, while lung representations improve the Harrell C-index for NLST lung-cancer mortality by 14.2\%. The global representation reaches INSPECT Recall@10 of 33.09\% and 32.04\% for text-to-image and image-to-text retrieval, respectively. Across organ-related tasks, anatomically matched representations provide stronger task-relevant signal, while the global representation retains broad utility. OrganLens offers a scalable approach to organ-specific CT representation learning with a shared encoder. More broadly, it provides the medical research community with a reusable framework for studying organ-specific disease across cohorts and clinical endpoints.

Figures

Figures reproduced from arXiv: 2607.25164 by Anqi Li, Hanwen Xu, Sadeer Al-Kindi, Wei Qiu, Zhixuan Ge.

Figure 1
Figure 1. Figure 1: Overview of OrganLens. (a) Organ-identity conditioning, organ-guided cropping, organ-specific distillation, and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Organ-specific abnormality detection on CT-RATE and RAD-ChestCT. (a) AUROC ranks of the 11 organ-specific [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Organ-specific prognostic evaluation on INSPECT and NLST. (a,b) Test Harrell C-indices for OrganLens, CT-pretrained [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Longitudinal NLST assessment using OrganLens. (a–d) Mean anatomically matched Cox log-hazard scores at T0–T2 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Image–text retrieval on CT-RATE (a,b) and INSPECT [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pretraining-component and representation-pooling [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] 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

58 extracted references · 8 canonical work pages · 1 internal anchor

  1. [1]

    Xuguang Bai, Mingxuan Liu, Tongxi Song, Yifei Chen, Hongjia Yang, Kasidit An- mahapong, Zihan Li, Ying Zhou, and Qiyuan Tian. 2026. EXACT: An Explainable Anomaly-Aware Vision Foundation Model for Analysis of 3D Chest CT.arXiv preprint arXiv:2604.24146(2026)

    Pith/arXiv arXiv 2026

  2. [2]

    Louis Blankemeier, Ashwin Kumar, Joseph Paul Cohen, Jiaming Liu, Longchao Liu, Dave Van Veen, Syed Jamal Safdar Gardezi, Hongkun Yu, Magdalini Paschali, Zhihong Chen, et al. 2026. Merlin: a computed tomography vision–language foundation model and dataset.Nature652, 8112 (2026), 1318–1328

    2026

  3. [3]

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258(2021)

    Pith/arXiv arXiv 2021

  4. [4]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging Prop- erties in Self-Supervised Vision Transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision. 9650–9660. https://openaccess.thecvf.com/content/ICCV2021/html/Caron_Emerging_ Properties_in_Self-Supervis...

    2021

  5. [5]

    Cris Claessens, Christiaan Viviers, Giacomo D’Amicantonio, Egor Bondarev, and Fons van der Sommen. 2025. Scaling Self-Supervised and Cross-Modal Pretraining for Volumetric CT Transformers.arXiv preprint arXiv:2511.17209(2025)

    arXiv 2025

  6. [6]

    Lavsen Dahal, Yubraj Bhandari, Geoffrey Rubin, and Joseph Y. Lo. 2026. ORACLE- CT: Anatomy-Aware Support Pooling for CT Classification.arXiv preprint arXiv:2606.05460(2026). doi:10.48550/arXiv.2606.05460

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.05460 2026

  7. [7]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInterna- tional Conference on Learning Representations. https:...

    2021

  8. [8]

    Rachel Lea Draelos, David Dov, Maciej A Mazurowski, Joseph Y Lo, Ricardo Henao, Geoffrey D Rubin, and Lawrence Carin. 2021. Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes.Medical image analysis67 (2021), 101857

    2021

  9. [9]

    Lo, Ricardo Henao, Geoffrey D

    Rachel Lea Draelos, David Dov, Maciej A Mazurowski, Joseph Y. Lo, Ricardo Henao, Geoffrey D. Rubin, and Lawrence Carin. 2022. RAD-ChestCT: Radiology Reports and CT Imaging Dataset. doi:10.5281/zenodo.6406114

    doi:10.5281/zenodo.6406114 2022

  10. [10]

    Yuxin Du, Fan Bai, Tiejun Huang, and Bo Zhao. 2024. SegVol: Universal and Inter- active Volumetric Medical Image Segmentation.Advances in Neural Information Processing Systems37 (2024), 110746–110783

    2024

  11. [11]

    Zebin Gao, Guoxun Zhang, Hengrui Liang, et al. 2026. A lung CT vision founda- tion model facilitating disease diagnosis and medical imaging.Nature Communi- cations17, 1 (2026), 35. doi:10.1038/s41467-025-66620-z

    work page doi:10.1038/s41467-025-66620-z 2026

  12. [12]

    Ioannis Gatopoulos, Nicolas Känzig, Sebastian Otálora, and Fei Tang. 2026. Coral- Bay: A Self-Supervised CT Foundation Model.arXiv preprint arXiv:2606.03888 (2026). https://arxiv.org/abs/2606.03888

    Pith/arXiv arXiv 2026

  13. [13]

    Ibrahim Ethem Hamamci, Sezgin Er, Chenyu Wang, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Omer Faruk Durugol, Benjamin Hou, Suprosanna Shit, et al. 2026. Generalist Foundation Models from a Multimodal Dataset for 3D Computed Tomography.Nature Biomedical Engineering(2026). doi:10.1038/s41551-025-01599-y

    doi:10.1038/s41551-025-01599-y 2026

  14. [14]

    Mohammad Reza Hosseinzadeh Taher, Junpyo Hong, Ravi Soni, and Gopal Avinash. 2025. Learning Foundation Models from Multi-organ Medical Images by Capturing Consistency and Diversity of Anatomical Structures. InInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 294–304

    2025

  15. [15]

    Kexin Huang, Jaan Altosaar, and Rajesh Ranganath. 2019. ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission.arXiv:1904.05342(2019)

    Pith/arXiv arXiv 2019

  16. [16]

    Shih-Cheng Huang, Zepeng Huo, Ethan Steinberg, Chia-Chun Chiang, Cur- tis Langlotz, Matthew Lungren, Serena Yeung, Nigam Shah, and Jason Fries

  17. [17]

    Shih-Cheng Huang, Liyue Shen, Matthew P Lungren, and Serena Yeung. 2021. GLoRIA: A Multimodal Global-Local Representation Learning Framework for Label-Efficient Medical Image Recognition. InProceedings of the IEEE/CVF inter- national conference on computer vision. 3942–3951

    2021

  18. [18]

    Junxian Li, Yuchen Xing, Ximin Gao, Zhaoxiang Ye, Meng Wang, and Fengju Song. 2026. A Self-Supervised Foundation Model Based on Three-Dimensional Chest CT Scans for Lung Cancer Diagnosis and Prognosis Prediction.Radiology: Imaging Cancer8, 2 (2026), e250360. doi:10.1148/rycan.250360

    work page doi:10.1148/rycan.250360 2026

  19. [19]

    Baciak, and Xiaofeng Yang

    Yuheng Li, Yuan Gao, Haoyu Dong, Yuxiang Lai, Shansong Wang, Mojtaba Safari, James E. Baciak, and Xiaofeng Yang. 2026. Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining.arXiv preprint arXiv:2605.21906(2026). https://arxiv.org/abs/2605.21906

    Pith/arXiv arXiv 2026

  20. [20]

    Jingyang Lin, Yingda Xia, Jianpeng Zhang, Ke Yan, Le Lu, Jiebo Luo, and Ling Zhang. 2024. CT-GLIP: 3D Grounded Language–Image Pretraining with CT Scans and Radiology Reports for Full-Body Scenarios.arXiv preprint arXiv:2404.15272 (2024)

    arXiv 2024

  21. [21]

    Jie Liu, Yixiao Zhang, Kang Wang, Mehmet Can Yavuz, Xiaoxi Chen, Yixuan Yuan, Haoliang Li, Yang Yang, Alan Yuille, Yucheng Tang, and Zongwei Zhou

  22. [22]

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. 2024. Segment anything in medical images.Nature communications15, 1 (2024), 654

    2024

  23. [23]

    Léo Machado, Léo Alberge, Hélène Philippe, et al . 2025. A promptable CT foundation model for solid tumor evaluation.npj Precision Oncology9, 1 (2025),

    2025

  24. [24]

    Shahin, Yukun Zhou, et al

    Niccolò McConnell, Pardeep Vasudev, Daisuke Yamada, Daryl Cheng, Mehran Azimbagirad, John McCabe, Shahab Aslani, Ahmed H. Shahin, Yukun Zhou, et al

  25. [25]

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. 2016. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. InFourth International Conference on 3D Vision, 3DV 2016, Stanford, CA, USA, October 25-28, 2016. IEEE Computer Society, 565–571. doi:10.1109/3DV.2016.79

    doi:10.1109/3dv.2016.79 2016

  26. [26]

    Amgad Muneer, Eman Showkatian, Yuliya Kitsel, et al. 2026. CT-Based Deep Foundation Model for Predicting Immune Checkpoint Inhibitor-Induced Pneu- monitis Risk in Lung Cancer.medRxiv(2026). doi:10.64898/2026.04.21.26351428

    work page doi:10.64898/2026.04.21.26351428 2026

  27. [27]

    National Lung Screening Trial Research Team. 2011. The National Lung Screening Trial: Overview and Study Design.Radiology258, 1 (jan 2011), 243–253. doi:10. 1148/radiol.10091808

    2011

  28. [28]

    Xuefeng Ni, Linshan Wu, Jiaxin Zhuang, Qiong Wang, Mingxiang Wu, Varut Vardhanabhuti, Lihai Zhang, Hanyu Gao, and Hao Chen. 2026. MG-3D: Multi- grained Knowledge-Enhanced Vision–Language Pre-training for 3D Medical Image Analysis.Medical Image Analysis111 (2026), 104027. doi:10.1016/j.media. 2026.104027

    arXiv 2026

  29. [29]

    Seok-Hwan Oh, Myeong-Gee Kim, Guil Jung, Hyeon-Jik Lee, Young-Min Kim, Sang-Yun Kim, Hyuksool Kwon, and Hyeon-Min Bae. 2026. Anatomy-aware Representation Learning for Medical Ultrasound. InThe Fourteenth Interna- tional Conference on Learning Representations. https://openreview.net/forum?id= 5ThIWuDkEf

    2026

  30. [30]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2024. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research(2024). https://openreview.net/forum? id=a68SUt6zFt

    2024

  31. [31]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings of the 38th Inter- national Conference on Machine Learning (Proceedings of Machi...

    2021

  32. [32]

    Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. 2019. Spreading Vectors for Similarity Search. InInternational Conference on Learning Representations. https://openreview.net/forum?id=SkGuG2R5tm

    2019

  33. [33]

    Zhongyi Shui, Jianpeng Zhang, Weiwei Cao, Sinuo Wang, Ruizhe Guo, Le Lu, Lin Yang, Xianghua Ye, Tingbo Liang, Qi Zhang, and Ling Zhang. 2025. Large-scale and Fine-grained Vision–Language Pre-training for Enhanced CT Image Under- standing. InThe Thirteenth International Conference on Learning Representations

    2025

  34. [34]

    Sifan Song, Siyeop Yoon, Pengfei Jin, Sekeun Kim, Matthew Tivnan, Yujin Oh, Runqi Meng, Ling Chen, Zhiliang Lyu, Dufan Wu, et al. 2025. OWT: A Founda- tional Organ-Wise Tokenization Framework for Medical Imaging.arXiv preprint arXiv:2505.04899(2025)

    arXiv 2025

  35. [35]

    Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang

  36. [36]

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. 2022. Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20730–20740

    2022

  37. [37]

    Jiayi Wang, Hadrien Reynaud, Ibrahim Ethem Hamamci, Sezgin Er, Suprosanna Shit, Bjoern Menze, and Bernhard Kainz. 2026. SigVLP: Sigmoid Volume– Language Pre-Training for Self-Supervised CT-Volume Adaptive Representation Learning.arXiv preprint arXiv:2602.21735(2026)

    arXiv 2026

  38. [38]

    Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. 2023. TotalSegmentator: robust segmentation of 104 anatomic KDD ’27 AI4Sciences, August 2027, San Jose, CA, USA Ge et al. structures in CT images.Radiology: Artificial Intelligence5, 5 (2...

    2023

  39. [39]

    Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2023. MedKLIP: Medical Knowledge Enhanced Language-Image Pre-Training for X-Ray Diagnosis. InProceedings of the IEEE/CVF international conference on computer vision. 21372–21383

    2023

  40. [40]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Representing part-whole hierarchies in foundation models by learning localizability composability and decomposability from anatomy via self supervi- sion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11269–11281

  41. [41]

    Qiangqiang Wu, Grace McIlvain, Zhou Yu, and Junhao Wen. 2026. Pan-FM: A Pan- Organ Foundation Model with Saliency-Guided Masking for Missing Robustness. arXiv preprint arXiv:2605.07055(2026)

    Pith/arXiv arXiv 2026

  42. [42]

    Yutong Xie, Jianpeng Zhang, Yong Xia, and Qi Wu. 2022. UniMiSS: Univer- sal Medical Self-Supervised Learning via Breaking Dimensionality Barrier. In European Conference on Computer Vision. Springer, 558–575

    2022

  43. [43]

    Peter Musi, Jeffrey Keenan, Maziar Khorsandi, Bassel Al-Alao, Ioan- nis Dimarakis, Hamid Chalian, Yiing Lin, Daniel Fishbein, Jay Pal, Sheng Wang, and Shin Lin

    Hanwen Xu, Addie Woicik, Sanaz Asadian, Junbo Shen, Zhengyan Zhang, Ali Nabipoor, J. Peter Musi, Jeffrey Keenan, Maziar Khorsandi, Bassel Al-Alao, Ioan- nis Dimarakis, Hamid Chalian, Yiing Lin, Daniel Fishbein, Jay Pal, Sheng Wang, and Shin Lin. 2025. A Cardiac-Specific CT Foundation Model for Heart Trans- plantation.medRxiv(2025). doi:10.1101/2025.08.14.25333618

    work page doi:10.1101/2025.08.14.25333618 2025

  44. [44]

    Kohei Yamamoto and Tomohiro Kikuchi. 2026. TotalFM: An Organ-Separated 3D-CT Foundation Model Leveraging Large-Scale Routine Clinical Radiology Data.arXiv preprint arXiv:2601.00260(2026)

    Pith/arXiv arXiv 2026

  45. [45]

    Linshan Wu, Jiaxin Zhuang, and Hao Chen. 2024. VoCo: A Simple-Yet-Effective Volume Contrastive Learning Framework for 3D Medical Image Analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 22873–22882

    2024

  46. [46]

    Guoliang You, Hongming Li, Yuanwang Zhang, and Yong Fan. 2026. Learn- ing Anatomy-Grounded CT Vision–Language Representations with Organ- Hierarchical Report Knowledge.arXiv preprint arXiv:2607.10953(2026)

    Pith/arXiv arXiv 2026

  47. [47]

    Huihui Yu and Qun Dai. 2026. Anatomy-Guided Self-Supervised Distillation Learning for Medical Image Analysis.IEEE Transactions on Medical Imaging45, 5 (2026), 1787–1801. doi:10.1109/TMI.2026.3680920

    arXiv 2026

  48. [48]

    Ke Yu, Li Sun, Junxiang Chen, Maxwell Reynolds, Tigmanshu Chaudhary, and Kayhan Batmanghelich. 2024. DrasCLR: A self-supervised framework of learning disease-related and anatomy-specific representation for 3D lung CT images. Medical Image Analysis92 (2024), 103062

    2024

  49. [49]

    Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Weidi Xie, and Yanfeng Wang. 2023. Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications14, 1 (2023), 4542

    2023

  50. [50]

    Jing Yang, Du Cai, Junwei Liu, Zhenfeng Zhuang, Yibin Zhao, Feng-ao Wang, Chenghang Li, Chuling Hu, Baowen Gai, Yiping Chen, Yixue Li, Liansheng Wang, Feng Gao, and Xiaojian Wu. 2025. CRCFound: A Colorectal Cancer CT Image Foundation Model Based on Self-Supervised Learning.Advanced Science12, 41 (2025), e07339. doi:10.1002/advs.202407339

    work page doi:10.1002/advs.202407339 2025

  51. [51]

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. 2022. iBOT: Image BERT Pre-Training with Online Tokenizer. In International Conference on Learning Representations. https://openreview.net/ forum?id=ydopy-e6Dg

    2022

  52. [52]

    Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. 2021. Models Genesis.Medical image analysis67 (2021), 101840

    2021

  53. [53]

    Frontera, Arjun V

    Weicheng Zhu, Haoxu Huang, Huanze Tang, Rushabh Musthyala, Boyang Yu, Long Chen, Emilio Vega, Thomas O’Donnell, Reya Hayek, Lindsey Kuohn, Seena Dehkharghani, Jennifer A. Frontera, Arjun V. Masurkar, Kara Melmed, and Narges Razavian. 2026. 3D Foundation Model for Generalizable Disease Detection in Head Computed Tomography.Nature Biomedical Engineering(apr...

    arXiv 2026

  54. [55]

    Hui Zhao, Ruipeng Zhang, Zhiyu Wang, Yifeng Gu, Shengyuan Xu, Sheng Wang, and Yuehua Li. 2026. BoneCoT: multicentre validation of a whole-body skele- ton foundation model for bone metastases guided by clinician-derived chain of thought.Nature Biomedical Engineering(jul 2026). doi:10.1038/s41551-026-01736- 1

    work page doi:10.1038/s41551-026-01736- 2026

  55. [121]

    doi:10.1038/s41698-025-00903-y

    doi:10.1038/s41698-025-00903-y

  56. [2023]

    InAdvances in Neural Information Processing Sys- tems, Vol

    INSPECT: A Multimodal Dataset for Patient Outcome Prediction of Pulmonary Embolisms. InAdvances in Neural Information Processing Sys- tems, Vol. 36. https://proceedings.neurips.cc/paper_files/paper/2023/hash/ 39736af1b9d87a1fddad9f84a6bcf64c-Abstract-Datasets_and_Benchmarks.html

    2023

  57. [2024]

    doi:10.1016/j.media.2024.103226

    Universal and extensible language–vision models for organ segmentation and tumor detection from abdominal computed tomography.Medical Image Analysis97 (2024), 103226. doi:10.1016/j.media.2024.103226

    arXiv 2024

  58. [2026]

    doi:10.1038/s43856-025-01328-1

    A computationally frugal, open-source chest CT foundation model for thoracic disease detection in lung cancer screening programmes.Communications Medicine6 (2026), 83. doi:10.1038/s43856-025-01328-1

    work page doi:10.1038/s43856-025-01328-1 2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.