SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training

Fenghe Tang; Haoran Lai; Qingsong Yao; Rongsheng Wang; Rui Yan; Shaohua Kevin Zhou; Xiaodong Tao; Xu Zhang; Zhen Huang; Zhiyang He

REVIEW 3 major objections 4 minor 36 references

SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training

T0 review · 3 major / 4 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 SimCroP pre-trains a chest CT encoder by aligning each report sentence to its most similar image patches, improving classification and segmentation over prior medical vision-language methods.

desk verdict A plausible and internally coherent Med-VLP recipe for chest CT whose claimed SOTA margins currently rest on re-implemented baselines with no reported details, single runs, no error bars, and one cited method omitted. read the letter →

arxiv 2509.08311 v1 pith:4HATOXLT submitted 2025-09-10 cs.CV

Rongsheng Wang , Fenghe Tang , Qingsong Yao , Rui Yan , Xu Zhang , Zhen Huang , Haoran Lai , Zhiyang He

show 3 more authors

Xiaodong Tao Zihang Jiang Shaohua Kevin Zhou

This is my paper

classification cs.CV

keywords medicalvision-languagepre-trainingchestCTsimilarity-drivenalignmentcross-granularityfusionmaskedautoencoderradiologyreportrepresentationlearning3Dimaging

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

SimCroP is a pre-training recipe for chest CT volumes paired with radiology reports. Its central idea is that each descriptive sentence in the Finding section should be matched to the specific image subregion it describes, and that this can be learned without any manual grounding labels by pulling each sentence embedding toward its top-K most similar patches. On top of that alignment, SimCroP also reconstructs masked image patches and masked report words, fusing whole-volume and patch-level features during the report reconstruction. The paper validates the recipe on linear-probe classification and fine-tuning segmentation across five public datasets and reports consistent gains over prior self-supervised and vision-language methods, e.g., 81.0 vs 79.4 AUC at 1% labeled data on CT-RATE and 93.5 vs 93.1 Dice at 10% data on LUNA16. If these results hold, the approach shows how unstructured radiology text can be used as a free source of fine-grained supervision for 3D medical image encoders.

What carries the argument

Similarity-driven Alignment (SA) is the load-bearing mechanism: for each sentence in the Finding section, the model computes dot-product similarity between the sentence embedding and every unmasked image patch, takes the top-K patches (K=64, about 10% of the radiograph area), aggregates their visual features by global average pooling, and pulls that aggregate toward the sentence embedding with a symmetric InfoNCE loss. This teaches the vision encoder, without any manual grounding labels, to associate descriptive sentences with the subregions that actually contain the described findings. Around this core, SimCroP stacks multi-modal masked modeling (MAE-style 75% patch masking with MSE reconst

What would settle it

Run the exact linear-probe protocol on CT-RATE's 1% split, once with SimCroP and once with the fVLM and M3AE checkpoints released by their original authors; if SimCroP no longer beats them by the reported margins (e.g., 81.0 vs 79.4 AUC), the central comparison collapses. The same applies to the LUNA16 10% Dice comparison.

Watch

Extended reading notes

Core claim

The paper claims that sentence-to-subregion alignment is the missing piece in medical vision-language pre-training for sparse 3D radiographs. Rather than contrasting a whole radiology report against a whole volume, SimCroP splits the Finding section into sentences, computes dot-product similarity between each sentence embedding and every unmasked volumetric patch, selects the top-K patches, and applies a symmetric InfoNCE loss between the pooled patch aggregate and the sentence. The reported result is that this mechanism, combined with masked image modeling and cross-granularity report reconstruction, produces chest CT representations that outperform both masked-modeling-only methods and con

Load-bearing premise

The gains are measured against baselines the authors re-implemented 'under the same data settings'; if those re-implementations are not faithful to the original methods, the reported lead over them may be an artifact rather than a property of SimCroP.

Editorial extensions

If this is right

With only 1% labeled CT-RATE data, the pre-trained encoder reaches 81.0 AUC, a 1.6-point margin over the strongest Med-VLP baseline, which would matter where annotations are scarce.
The representation transfers beyond chest classification: 93.5 Dice at 10% LUNA16 data and 80.7 on BTCV show gains on nodule and abdominal organ segmentation.
Ablations indicate the similarity-driven alignment contributes more than 1% AUC and 1.1 Dice over masked modeling alone, suggesting sentence-level supervision supplies a fine-grained signal that reconstruction loss lacks.
Including both instance-level and word-patch features helps more than either alone, with instance-level features especially important for lung segmentation because macro-anatomy dominates the volume.
Because pre-training uses only paired CT-reports without manual grounding annotations, the recipe can scale to larger unlabeled radiology corpora.

Reading between the lines

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

If top-K selection reliably highlights lesion-bearing patches, as the visualization suggests, the same similarity scores could be converted into a weakly supervised localization or segmentation signal without box or pixel labels; the paper does not train such a head or quantify localization accuracy.
The method depends on the Finding section's sentence structure, so a testable extension is whether paragraph-aware weighting or harder negatives across patients improves alignment beyond symmetric in-batch InfoNCE.
Because the paper re-implements most baselines rather than using their official checkpoints, an external re-run with the original releases would separate genuine method gains from implementation details; this is a verification step the paper does not itself carry out.
The same recipe may transfer to other sparse 3D modalities, such as MRI volumes with paired reports, since the core design does not rely on CT-specific Hounsfield-unit normalization; the paper only demonstrates CT.

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.

Desk Editor's Note

read the letter

SimCroP is a coherent Med-VLP recipe for chest CT: masked multimodal modeling, top-K sentence-to-patch contrastive alignment, and cross-granularity fusion (instance-level GAP plus word-patch cross-attention) for masked report reconstruction. The combination is novel enough, even if each ingredient is derived from M3AE, CLIP-style contrastive learning, and fVLM/BIUD-style fine-grained alignment. What the paper does well: the method is clearly described, the ablation in Table 3 shows each component contributes (SA gives about 1% AUC and 1.1% Dice), the top-K visualization is a nice sanity check, and the conclusion honestly states the method's zero-shot limitation. That matters; the authors are not overclaiming the scope.

The soft spot is the evidence behind the headline 'outperforms SOTA'. The comparisons in Tables 1 and 2 are against re-implementations of M3AE, CT-CLIP, MRM, and fVLM. Section 3 says 'same data settings,' but that controls pre-training data, not method fidelity. No hyperparameters, masking schedules, epochs, seeds, or evaluation protocols are reported, and no baseline code or checkpoints are available. One cited SOTA method, MG-3D, is not in the tables. Margins are modest—1.6 AUC at 1% CT-RATE, 0.4 Dice at 10% LUNA16—so a slightly weak re-implementation of fVLM or MRM could easily explain the gap. There are also no error bars or significance tests; single runs. And the paper uses test sets as validation sets when no official validation exists, and the top-K and masking-ratio hyperparameters in Fig. 3 were selected on the same downstream benchmarks used for final claims. None of this is misconduct, but it means the SOTA claim is provisional, not established.

Who is this for: people building CT pre-training pipelines who want a concrete recipe with component-level ablations. The method is worth trying, and the code URL suggests it may be reproducible. But a serious editor should send this to peer review only with the expectation that the authors supply full baseline implementation details, error bars, and a head-to-head with MG-3D. If those come through, the paper is a useful contribution.

Referee Report

3 major / 4 minor

Summary. The paper proposes SimCroP, a medical vision-language pre-training method for 3D chest CT volumes. The framework combines masked image modeling, similarity-driven sentence-to-top-K-patch alignment, and a cross-granularity fusion module that merges instance-level and word-patch-level features for masked report reconstruction. Pre-trained on CT-RATE and evaluated with linear-probe classification and fine-tuning segmentation on five public datasets, the method reports consistent improvements over several Med-SSL and Med-VLP baselines, with the largest margins being 1.6 AUC on CT-RATE 1% classification and 0.4 Dice on LUNA16 10% segmentation. The authors provide internal ablations supporting each component and a qualitative visualization of the learned patch-sentence alignment.

Significance. If the reported results hold, SimCroP would represent a meaningful advance in fine-grained Med-VLP for sparse 3D radiographs, with practical value for label-efficient classification and segmentation. The design is plausible: combining masked modeling with similarity-driven alignment is a reasonable way to exploit sentence-level supervision without spatial annotations, and the internal ablations in Table 3 support the contribution of each module. The paper also uses public datasets and states that code and models will be released. However, the comparisons are only as strong as the faithfulness of the re-implemented baselines and the statistical reliability of the single-run numbers, both of which are currently not verifiable from the manuscript.

major comments (3)

[§3, Baselines; Tables 1–2] The central claim of outperforming existing methods is measured exclusively against re-implemented baselines (M3AE, CT-CLIP, MRM, fVLM) plus official M3D weights. The manuscript does not report the baselines' hyperparameters, masking schedules, optimizer settings, training epochs, random seeds, or evaluation protocols, and no code/checkpoints are provided in the preprint. Since methods such as CT-CLIP and fVLM have specific contrastive/fine-grained designs that are easy to weaken inadvertently, the reported margins (e.g., 1.6 AUC, 0.4 Dice) may reflect implementation differences rather than genuine improvements. The authors should release the re-implementation code and full training/evaluation configurations, or use official checkpoints under a common protocol.
[§3, Implementation; Tables 1–3] All experiments appear to be single runs with no error bars or significance tests. Moreover, the paper states: 'Unless official validation set is defined, testing sets serve as validation sets for all experimental configurations.' This means that the hyperparameter choices in Fig. 3 (Top-K=64, report masking ratio=75%) and the final numbers in Tables 1–2 are selected on the same test sets used for the headline claims. This introduces optimistic bias and prevents assessing whether the differences are statistically meaningful. The authors should report mean±std over multiple seeds and use held-out validation splits for model/hyperparameter selection.
[§1, Related work; §3, Baselines] MG-3D [19] is cited in the introduction as a relevant recent method ('incorporates both intra-patient cross-modal semantic consistency and inter-patient semantic correlations'), but it is absent from the comparisons in Tables 1–2. Given the paper's claim to outperform 'cutting-edge' and 'state-of-the-art' methods, the omission of a directly related cited method weakens the comparison. The authors should either include MG-3D in the experiments or justify its exclusion (e.g., no public weights) clearly.

minor comments (4)

[§2.2 heading] The heading contains a typo: 'Simlilarity-driven Alignment' should be 'Similarity-driven Alignment'.
[§2.3, last paragraph] 'The decoding procedure will be introduced in secion 2.3' should be 'section 2.3'.
[Eq. (3)] The notation Sim^{l,K}(T^l_sent, I^s_u) = TopK_{0≤s<N} Sim^{l,s}(...) is ambiguous: it should clarify that the right-hand side selects the K patch indices with largest similarity, and how the resulting set is used in the subsequent GAP operation.
[Fig. 2] The figure showing per-label AUC on RadChestCT lacks axis labels and a legend for the metric. Please add a y-axis label (e.g., AUC) and specify the evaluation setting in the caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: SimCroP's pre-training objectives and downstream evaluations are independent; minor self-citations are not load-bearing.

full rationale

The paper proposes a pre-training framework and validates it on external public benchmarks. The three loss terms (L_MIM, L_align, L_MLM) are defined from the model's inputs and standard objectives; none is defined in terms of the downstream classification/segmentation metrics used for evaluation. The similarity-driven alignment uses the model's own similarity to select top-K patches and then pulls sentence features toward those patches; this is a self-supervised training objective, not a prediction forced by construction. Hyperparameters Top-K=64 and report masking ratio=75% are chosen via ablations on the same datasets later used for evaluation (Fig. 3), which is a model-selection concern, but the reported numbers are not equal to the fit term by construction. The paper cites several prior works by the same group (e.g., [27], [28], [31]), but none of these citations supplies a load-bearing premise; the architecture uses external components (MAE initialization, CXR-BERT) and is compared against external baselines. Reproducibility concerns about re-implemented baselines and test-set-as-validation are correctness/verifiability issues, not circularity. The conclusion explicitly acknowledges a limitation (lack of instance-level cross-modal alignment hinders zero-shot performance), which further indicates that the claims are not being protected by definition.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the CT-RATE paired dataset, on public pretrained encoders, and on hyperparameters such as Top-K, mask ratio, loss weights, and temperature. Several hyperparameters were tuned on the same downstream benchmarks used for the final tables, so a portion of the reported gain is fitted. No new physical or conceptual entities beyond the learned network modules are introduced.

free parameters (4)

Top-K patches per sentence = 64
Number of most similar patches selected per sentence in similarity-driven alignment; tuned via ablation on RadChestCT and LUNA16 in Fig. 3.
Report masking ratio gamma = 0.75
Proportion of report tokens masked; selected via ablation in Fig. 3.
Loss weights lambda1, lambda2 = 1, 1
Weights for alignment and masked-report losses, set empirically to balance objectives per Sec. 2.3.
Temperature tau = 0.07
Temperature in the InfoNCE loss in Eq. 4, fixed following common practice.

assumptions (5)

domain assumption Top-K similarity selection identifies the subregions described by each sentence without spatial grounding
Central premise of similarity-driven alignment in Sec. 2.2; the paper provides no spatial grounding annotations and assumes cosine similarity is a valid surrogate.
domain assumption Pretrained CXR-BERT and MAE ImageNet-1K weights transfer useful knowledge to 3D chest CT encoders
Sec. 3 Implementation initializes the vision and text encoders from these pretrained weights, assuming cross-modality and cross-dimension transfer.
domain assumption CT-RATE report findings sentences correspond to localized visual subregions
Motivating assumption stated in Sec. 1 and used by the sentence alignment loss in Sec. 2.2.
domain assumption Re-implemented baselines faithfully reproduce the original methods
Sec. 3 Baselines notes M3D uses official weights and the other methods are re-implemented under the same data settings; the comparison's fairness depends on this.
standard math InfoNCE contrastive loss effectively aligns sentence and patch embeddings
Eq. 4 uses the standard InfoNCE objective, assumed to be a suitable training signal for patch-sentence alignment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training." pith.science (2026). https://pith.science/paper/4HATOXLT

@misc{pith2026250908311,
  author       = {Pith},
  title        = {Pith review of: SimCroP: Radiograph Representation Learning with Similarity-driven Cross-granularity Pre-training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HATOXLT}},
  note         = {Machine review of arXiv:2509.08311}
}

read the original abstract

Medical vision-language pre-training shows great potential in learning representative features from massive paired radiographs and reports. However, in computed tomography (CT) scans, the distribution of lesions which contain intricate structures is characterized by spatial sparsity. Besides, the complex and implicit relationships between different pathological descriptions in each sentence of the report and their corresponding sub-regions in radiographs pose additional challenges. In this paper, we propose a Similarity-Driven Cross-Granularity Pre-training (SimCroP) framework on chest CTs, which combines similarity-driven alignment and cross-granularity fusion to improve radiograph interpretation. We first leverage multi-modal masked modeling to optimize the encoder for understanding precise low-level semantics from radiographs. Then, similarity-driven alignment is designed to pre-train the encoder to adaptively select and align the correct patches corresponding to each sentence in reports. The cross-granularity fusion module integrates multimodal information across instance level and word-patch level, which helps the model better capture key pathology structures in sparse radiographs, resulting in improved performance for multi-scale downstream tasks. SimCroP is pre-trained on a large-scale paired CT-reports dataset and validated on image classification and segmentation tasks across five public datasets. Experimental results demonstrate that SimCroP outperforms both cutting-edge medical self-supervised learning methods and medical vision-language pre-training methods. Codes and models are available at https://github.com/ToniChopp/SimCroP.

Figures

Figures reproduced from arXiv: 2509.08311 by the authors.

**Figure 3.** Ablation on Top-K and report masking ratio selection. assess cross-domain transferability of learned chest CT representations to abdominal imaging, we employ the BTCV [16] dataset under its prescribed organ segmentation protocol with official data splits. Unless official validation set is defined, testing sets serve as validation sets for all experimental configurations. Implementation. We resample the volume to 1.… view at source ↗

**Figure 4.** Visualization of top-K patches correlated to descriptive sentences. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

[19]

Ni, X., Wu, L., Zhuang, J., Wang, Q., et al.: Mg-3d: Multi-grained knowledge- enhanced3dmedicalvision-languagepre-training.arXivpreprintarXiv:2412.05876 (2024)

arXiv 2024
[1]

arXiv preprint arXiv:2404.00578 (2024)

Bai, F., Du, Y., Huang, T., Meng, M.Q.H., Zhao, B.: M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578 (2024)

arXiv 2024
[2]

In: ECCV

Boecking, B., Usuyama, N., Bannur, S., Castro, D.C., et al.: Making the most of text semantics to improve biomedical vision–language processing. In: ECCV. pp. 1–21 (2022)

work page 2022
[3]

In: CVPR

Cao, W., Zhang, J., Xia, Y., Mok, T.C., et al.: Bootstrapping chest ct image understanding by distilling knowledge from x-ray expert models. In: CVPR. pp. 11238–11247 (2024)

work page 2024
[4]

Medical image analysis79, 102444 (2022)

Chen, X., Wang, X., Zhang, K., Fung, K.M., Thai, T.C., Moore, K., Mannel, R.S., Liu, H., Zheng, B., Qiu, Y.: Recent advances and clinical applications of deep learning in medical image analysis. Medical image analysis79, 102444 (2022)

2022
[5]

In: MICCAI

Chen, Z., Du, Y., Hu, J., Liu, Y., Li, G., Wan, X., Chang, T.H.: Multi-modal masked autoencoders for medical vision-and-language pre-training. In: MICCAI. pp. 679–689. Springer (2022)

work page 2022
[6]

In: ICLR (2020)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)

work page 2020
[7]

Medical image analysis67, 101857 (2021)

Draelos,R.L.,Dov,D.,Mazurowski,M.A.,Lo,J.Y.,Henao,R.,Rubin,G.D.,Carin, L.: Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Medical image analysis67, 101857 (2021)

2021

Show all 36 references

[8]

CoRR (2024)

Hamamci, I.E., Er, S., Almas, F., Simsek, A.G., Esirgun, S.N., Dogan, I., Dasdelen, et al.: A foundation model utilizing chest ct volumes and radiology reports for supervised-level zero-shot detection of abnormalities. CoRR (2024)

2024
[9]

In: WACV

Hatamizadeh, A., Tang, Y., et al.: Unetr: Transformers for 3d medical image seg- mentation. In: WACV. pp. 574–584 (2022) 10 R. Wang, F. Tang et al

2022
[10]

In: CVPR

He, K., Chen, X., Xie, S., Li, Y., Doll’ar, P., Girshick, R.B.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 15979–15988 (2021)

2021
[11]

MedRxiv pp

He, X., Wang, S., Shi, S., Chu, X., Tang, J., Liu, X., Yan, C., Zhang, J., Ding, G.: Benchmarking deep learning models and automated model design for covid-19 detection with chest ct scans. MedRxiv pp. 2020–06 (2020)

2020
[12]

In: ICCV

Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: ICCV. pp. 3942–3951 (2021)

2021
[13]

Nature Communications15(1), 7620 (2024)

Huang, W., Li, C., Zhou, H.Y., Yang, H., Liu, J., et al.: Enhancing representation in radiography-reports foundation model: A granular alignment algorithm using masked contrastive learning. Nature Communications15(1), 7620 (2024)

2024
[14]

IJCARS19(5), 939–950 (2024)

Huang, Z., Li, H., Shao, S., Zhu, H., Hu, H., Cheng, Z., Wang, J., Kevin Zhou, S.: Pele scores: pelvic x-ray landmark detection with pelvis extraction and enhance- ment. IJCARS19(5), 939–950 (2024)

2024
[15]

JKS University Computer and Information Sciences37(3), 1–18 (2025)

Huang, Z., Zhou, X., He, X., Wei, Y., Yang, W., Wang, S., Sun, X., Li, H.: Case- mark: A hybrid model for robust anatomical landmark detection in multi-structure x-rays. JKS University Computer and Information Sciences37(3), 1–18 (2025)

2025
[16]

In: Proc

Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi- atlas labeling beyond the cranial vault–workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop. vol. 5, p. 12 (2015)

2015
[17]

In: CVPR

Li, Z., Yang, L.T., Ren, B., et al.: Mlip: Enhancing medical visual representation with divergence encoder and knowledge-guided contrastive learning. In: CVPR. pp. 11704–11714 (2024)

2024
[18]

Advances in NeurIPS32(2019)

Lu,J.,Batra,D.,Parikh,D.,Lee,S.:Vilbert:Pretrainingtask-agnosticvisiolinguis- tic representations for vision-and-language tasks. Advances in NeurIPS32(2019)

2019
[20]

ArXivabs/1807.03748(2018)

van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. ArXivabs/1807.03748(2018)

2018 arXiv
[21]

Irbm43(2), 87–92 (2022)

Pathak, Y., Shukla, P.K., Tiwari, A., Stalin, S., Singh, S.: Deep transfer learning based classification model for covid-19 disease. Irbm43(2), 87–92 (2022)

2022
[22]

In: ICML

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PMLR (2021)

2021
[23]

Medical image analysis42, 1–13 (2017)

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

2017
[24]

arXiv preprint arXiv:2304.13416 (2023)

Shao, S., Yuan, X., Huang, Z., et al.: Diffuseexpand: Expanding dataset for 2d med- ical image segmentation using diffusion models. arXiv preprint arXiv:2304.13416 (2023)

2023 arXiv
[25]

In: ICLR (2025)

Shui, Z., Zhang, J., Cao, W., et al.: Large-scale and fine-grained vision-language pre-training for enhanced ct image understanding. In: ICLR (2025)

2025
[26]

Medical Image Analysis p

Tang, F., Nian, B., Li, Y., et al.: Mambamim: Pre-training mamba with state space token interpolation and its application to medical image segmentation. Medical Image Analysis p. 103606 (2025)

2025
[27]

In: MICCAI

Tang, F., Xu, R., Yao, Q., et al.: Hyspark: Hybrid sparse masking for large scale medical image pre-training. In: MICCAI. pp. 330–340. Springer (2024)

2024
[28]

arXiv preprint arXiv:2502.08347 (2025) Similarity-driven Cross-granularity Pre-training 11

Tang, F., Yao, Q., et al.: Hi-end-mae: Hierarchical encoder-driven masked autoen- coders are stronger vision learners for medical image segmentation. arXiv preprint arXiv:2502.08347 (2025) Similarity-driven Cross-granularity Pre-training 11

2025 arXiv
[29]

Nature Biomedical Engineering6(12), 1399–1406 (2022)

Tiu, E., Talius, E., Patel, P., Langlotz, C.P., Ng, A.Y., Rajpurkar, P.: Expert-level detection of pathologies from unannotated chest x-ray images via self-supervised learning. Nature Biomedical Engineering6(12), 1399–1406 (2022)

2022
[30]

In: NeurIPS

Wang, F., Zhou, Y., et al.: Multi-granularity cross-modal alignment for general- ized medical visual representation learning. In: NeurIPS. vol. 35, pp. 33536–33549 (2022)

2022
[31]

arXiv preprint arXiv:2312.13316 (2023)

Wang, R., Yao, Q., Lai, H., He, Z., Tao, X., Jiang, Z., Zhou, S.: Ecamp: Entity-centered context-aware medical vision language pre-training. arXiv preprint arXiv:2312.13316 (2023)

2023 arXiv
[32]

In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

Wang, Z., Wu, Z., Agarwal, D., Sun, J.: MedCLIP: Contrastive learning from unpaired medical images and text. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 3876–3887 (Dec 2022)

2022
[33]

In: ICLR (2023)

Zhou,H.Y.,Lian,C.,Wang,L.,Yu,Y.:Advancingradiographrepresentationlearn- ing with masked record modeling. In: ICLR (2023)

2023
[34]

In: 2023 ISBI

Zhou, L., Liu, H., Bae, J., He, J., Samaras, D., Prasanna, P.: Self pre-training with masked autoencoders for medical image classification and segmentation. In: 2023 ISBI. pp. 1–6. IEEE (2023)

2023
[35]

Proceedings of the IEEE (2021)

Zhou, S.K., Greenspan, H., Davatzikos, C., Duncan, J.S., et al.: A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proceedings of the IEEE (2021)

2021
[36]

arXiv preprint arXiv:2412.06499 (2024)

Zhou, X., Huang, Z., Zhu, H., Yao, Q., Zhou, S.K.: Hybrid attention net- work: An efficient approach for anatomy-free landmark detection. arXiv preprint arXiv:2412.06499 (2024)

2024 arXiv

Pith tools

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

[1] [19]

Ni, X., Wu, L., Zhuang, J., Wang, Q., et al.: Mg-3d: Multi-grained knowledge- enhanced3dmedicalvision-languagepre-training.arXivpreprintarXiv:2412.05876 (2024)

arXiv 2024

[2] [1]

arXiv preprint arXiv:2404.00578 (2024)

Bai, F., Du, Y., Huang, T., Meng, M.Q.H., Zhao, B.: M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578 (2024)

arXiv 2024

[3] [2]

In: ECCV

Boecking, B., Usuyama, N., Bannur, S., Castro, D.C., et al.: Making the most of text semantics to improve biomedical vision–language processing. In: ECCV. pp. 1–21 (2022)

work page 2022

[4] [3]

In: CVPR

Cao, W., Zhang, J., Xia, Y., Mok, T.C., et al.: Bootstrapping chest ct image understanding by distilling knowledge from x-ray expert models. In: CVPR. pp. 11238–11247 (2024)

work page 2024

[5] [4]

Medical image analysis79, 102444 (2022)

Chen, X., Wang, X., Zhang, K., Fung, K.M., Thai, T.C., Moore, K., Mannel, R.S., Liu, H., Zheng, B., Qiu, Y.: Recent advances and clinical applications of deep learning in medical image analysis. Medical image analysis79, 102444 (2022)

2022

[6] [5]

In: MICCAI

Chen, Z., Du, Y., Hu, J., Liu, Y., Li, G., Wan, X., Chang, T.H.: Multi-modal masked autoencoders for medical vision-and-language pre-training. In: MICCAI. pp. 679–689. Springer (2022)

work page 2022

[7] [6]

In: ICLR (2020)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020)

work page 2020

[8] [7]

Medical image analysis67, 101857 (2021)

Draelos,R.L.,Dov,D.,Mazurowski,M.A.,Lo,J.Y.,Henao,R.,Rubin,G.D.,Carin, L.: Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Medical image analysis67, 101857 (2021)

2021