Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A three-stage recipe of self-supervised pretraining, finetuning, and test-time adaptation yields a cervical cancer screening system that generalizes across six external centers at 0.950 AUC.

desk verdict A large, prospectively validated cervical cytology system whose central external-generalization claim currently rests on an unstated and arithmetically questionable center-level split. read the letter →

arxiv 2502.09662 v1 pith:SYKQZ5ZJ submitted 2025-02-12 q-bio.QM cs.CVeess.IV

Hao Jiang , Cheng Jin , Huangjing Lin , Yanning Zhou , Xi Wang , Jiabo Ma , Li Ding , Jun Hou
show 10 more authors
Runsheng Liu Zhizhong Chai Luyang Luo Huijuan Shi Yinling Qian Qiong Wang Changzhong Li Anjia Han Ronald Cheong Kin Chan Hao Chen
This is my paper · ORCID
classification q-bio.QMcs.CVeess.IV
keywords cervicalcancerscreeningwhole-slideimagesself-supervisedpretrainingtest-timeadaptationDINOv2DeformableDETRgeneralizabilityCCS-127K
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

Smart-CCS, the system this paper introduces, is a claim about how to make AI cervical cancer screening work outside the hospitals where the model was trained. The proposed recipe is three sequential stages: self-supervised pretraining on a large multi-center cytology dataset, supervised finetuning of an abnormal-cell detector and a whole-slide classifier, and a test-time adaptation step that adjusts the model on the current clinical batch before predicting. On its own retrospective and prospective evaluations, the system reports 0.965 AUC over 11 internal test centers, 0.950 AUC over 6 external centers, and AUCs of 0.947, 0.924, and 0.986 at three prospective centers, with histology-confirmed diagnosis sensitivity of 0.943 on 738 biopsy-validated cases. The stakes are practical: if this holds, a single screening AI could be deployed at new hospitals without per-center retraining, which is the main reason prior cytology systems degrade in real-world use.

What carries the argument

The load-bearing object is the CCS-127K dataset (127,471 whole-slide images from 48 centers, 227 million patches, 104,979 cell-level annotations across six abnormal grades), because it is the resource that makes large-scale cytology pretraining possible. The mechanism is the three-stage Smart-CCS pipeline: self-supervised pretraining with a student-teacher vision transformer trained by masked reconstruction and prototype alignment; finetuning of a Deformable DETR abnormal-cell detector followed by a frozen-feature top-k slide classifier; and test-time adaptation with a mean-teacher update that aligns current-batch features and predictions to a class-wise prototype bank from retrospective data. Pretraining is the component the paper credits with domain-invariant representations, and adaptation is the component that closes the remaining gap at inference time.

What would settle it

Find the center identifiers behind the cohorts: if any of RC8-RC13 supplied unlabeled patches to the 112,062-slide pretraining cohort (or labeled slides to the 49,063-slide finetuning cohort), the reported 0.950 external AUC is inflated. A clean experiment is to retrain or re-run the pipeline with those six centers' data excluded from pretraining and finetuning and compare external AUC; the paper's own Base/w-P/w-P&A ablation provides the template for measuring the drop.

Watch

Extended reading notes

Core claim

On its own terms, this paper claims that generalizable cervical cancer screening is achievable by treating domain shift as a data problem rather than an architecture problem. It curates CCS-127K, 127,471 cytology whole-slide images from 48 centers with slide-level Bethesda grades and 104,979 abnormal-cell bounding-box annotations, then runs DINOv2-style self-supervised pretraining on 227 million patches so the feature extractor learns cytology structure (nuclei, cytoplasm, cell morphology) that is not tied to one scanner or stain. The pretrained encoder is frozen and used in a two-step screening pipeline: a Deformable DETR detector finds abnormal cells, and a top-k slide classifier aggregates cell features into a slide-level grade. Before inference on a new center, a test-time adaptation stage updates the classifier by aligning the current batch's features and predictions with a class-wise prototype bank built from retrospective data, using a mean-teacher loss. With that pipeline, the paper reports internal-test screening at 0.965 AUC (sensitivity 0.913), external testing at 0.950 AUC across six centers, prospective AUCs of 0.947, 0.924, and 0.986, and an ablation in which pretraining alone lifts external AUC from a 0.897-0.830 baseline range to 0.930-0.911 and the full pretraining-plus-adaptation system reaches 0.950.

Load-bearing premise

The load-bearing assumption is that the six external test centers were genuinely unseen by the model before evaluation, meaning their patches did not enter the self-supervised pretraining cohort or the finetuning cohort; the paper's center counts (83 center-roles described over a 48-center dataset) do not by themselves establish that this exclusion happened.

Editorial extensions

If this is right

  • If the external-test result is genuine, hospitals can adopt one screening model and expect roughly 0.950 AUC on their existing cytology WSIs without collecting new labels for their own scanners.
  • Pretraining scale is a direct lever: the paper's scaling experiments show WSI-level epithelial-abnormality accuracy rising by 10.34 percentage points as pretraining data grows to 100M patches, so building larger cytology corpora should transfer to other cytology tasks.
  • Test-time adaptation contributes about a 6.3-percentage-point AUC gain over the base pipeline on external centers, meaning inference-time adaptation is a usable replacement for per-center finetuning when labels are unavailable.
  • Histology-verified diagnosis sensitivity of 0.943 (and near 1.00 for HSIL+ lesions) implies the screening scores track actual cervical neoplasia rather than only the cytology labels used for training.
  • The two-step detection-plus-MIL design is not tightly coupled to one classifier head: S4MIL and TransMIL reach 0.969 and 0.966 AUC internally with the same features, so future improvements in slide aggregation can be dropped into the paradigm.

Reading between the lines

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

  • Editorial check on the independence claim: the study reports 39 pretraining centers, 27 finetuning centers, 11 internal-test centers, and 6 external-test centers, which sums to 83 center-roles over a 48-center pool; the paper never states that the six external centers (RC8-RC13) were excluded from pretraining, so the 0.950 external AUC should be read as conditional on that exclusion being true.
  • The Discussion's limitation paragraph says 'independent testing across 11 external centers,' while the Results describes six external centers (RC8-RC13); this internal inconsistency needs resolution before the generalization claim can be pinned down.
  • If the paradigm transfers as the paper's final discussion anticipates, the same pretraining-plus-adaptation recipe could be applied to urine, thyroid, and pleural-effusion cytology to build a multi-cancer cytology foundation model, but that extension is speculative because only cervical data are tested here.
  • The scaling-law trend suggests the remaining generalizability gap is a data-size problem more than an algorithm problem; adding more unlabeled cytology WSIs to pretraining is the cheapest testable lever for further gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript introduces Smart-CCS, a cervical cancer screening system that combines large-scale self-supervised pretraining on a newly curated multi-center cytology whole-slide image dataset (CCS-127K, 127,471 WSIs from 48 centers), finetuning of a two-step abnormal-cell-detection plus WSI-classification model, and test-time adaptation. The authors report a retrospective internal-test AUC of 0.965, an external-test AUC of 0.950 across six centers, and prospective AUCs of 0.947, 0.924, and 0.986 in three centers, with histology-confirmed diagnostic sensitivity. The paper claims that the pretraining-plus-adaptation paradigm generalizes across diverse clinical settings.

Significance. The dataset scale and the combination of self-supervised pretraining with test-time adaptation for cervical cytology are of considerable interest to computational pathology. The multi-center evaluation with confidence intervals and the prospective validation against histological ground truth are notable strengths. If the center-level independence of the external test set were conclusively established, the external-test result would be a meaningful demonstration of generalizability. The authors' candid discussion of limitations, including the smaller scale relative to histology foundation models, is creditworthy. However, the current manuscript does not provide sufficient evidence for the central independence claim, and the reported cohort arithmetic is internally inconsistent.

major comments (3)
  1. [Fig. 1(c) and Extended Data Table 14] The reported cohort sizes are arithmetically inconsistent with the dataset totals. The pretraining (112,062), finetuning (49,063), internal test (11,722), and external test (9,155) slide counts sum to 182,002, whereas the retrospective cohort contains 124,118 slides; the center counts (39+27+11+6=83) also exceed the 45 retrospective centers. Extended Data Table 14 lists RC8–RC13 as containing 12,056 WSIs, but only 9,155 are used for external testing, leaving 2,901 WSIs from those 'independent' centers unaccounted. Unless the authors provide an explicit center-level partition demonstrating that no RC8–RC13 slide entered pretraining or finetuning, the claim that the external test is independent is unverified. This is load-bearing for the central generalization claim.
  2. [Test-time adaptation with prototype alignment] The evaluation protocol adapts the model on the current test batch before prediction. The external-test AUC of 0.950 is therefore a transductive result, not a frozen-model generalization result. The paper should explicitly state that test-time adaptation is applied to the test batch and should present the frozen-model (w/ P) results as the inductive generalization reference; the abstract's phrasing 'across 6 independent test datasets' should be qualified accordingly.
  3. [Data availability] The manuscript states that source code will be released 'upon paper acceptance' and that only summary statistics of the private CCS-127K dataset are provided. In light of the cohort-arithmetic inconsistency above, a center-level study flowchart or data partition table listing, for each center, the number of slides used in pretraining, finetuning, internal testing, and external testing is necessary to verify the independence claim. Without this, the external-test results cannot be audited.
minor comments (5)
  1. [Discussion] The first paragraph of the Limitations section says 'independent testing across 11 external centers'; this should read '6 external centers' to match the abstract and Results.
  2. [Extended Data Table 4] 'ViT-Gaint' is a typo for 'ViT-Giant'.
  3. [Abstract] 'real-word practice' should be 'real-world practice'.
  4. [Methods, Evaluation metrics and setting] 'we tuned heyeparameters' is a typo for 'hyperparameters'.
  5. [Eq. (2)] The denominator notation 'Sim(zi·zj )' should be 'Sim(zi, zj)' for consistency with the definition immediately above.

Circularity Check

1 steps flagged · score 5.0 of 10

External 'independent' centers are not arithmetically disjoint from pretraining, so the 0.950 AUC is not verifiably an unseen-center prediction; prospective results keep partial independence.

  1. fitted input called prediction [Results, 'External testing for generalizability evaluation'; Fig. 1(c); Extended Data Table 14]
    "The retrospective cohort was used for model development and evaluation, and it was further divided into three subsets: pretraining, finetuning, and testing. ... an external test dataset with 9,155 samples from 6 independent centers."

    By the paper's own accounting the split cannot be disjoint: 112,062 + 49,063 + 11,722 + 9,155 = 182,002 WSIs, while the retrospective cohort is stated as 124,118 WSIs from 45 centers; the stage center counts (39+27+11+6=83) also exceed 45. Extended Data Table 14 gives RC8-RC13 a total of 12,056 WSIs, yet only 9,155 are used as the 'external' test; the other 2,901 WSIs from these six supposedly independent centers are assigned to no reported subset. Hence at least some of the pretraining cohort (39 centers; 39+6=45 is at least consistent with full overlap) can include the same centers as the external test.

full rationale

The method itself is not a circular derivation: DINOv2 pretraining, DDETR detection, MIL classification, and prototype-alignment test-time adaptation are all externally grounded components, and no load-bearing claim reduces to a self-citation chain. The central circularity concern is confined to the evaluation protocol. The paper claims the retrospective cohort was 'divided into three subsets' and that RC8-RC13 are '6 independent centers,' but the reported volumes (112,062 pretraining + 49,063 finetuning + 11,722 internal + 9,155 external = 182,002) exceed the 124,118-slide retrospective cohort, and the center entries (39+27+11+6=83) exceed the 45 retrospective centers. Extended Data Table 14 further shows 12,056 WSIs for RC8-RC13 while only 9,155 are externally tested, leaving 2,901 WSIs from these 'independent' centers unaccounted for in the reported test subset. Consequently, the external-generalization result may be partially fitted input called prediction: unlabeled patches from the same centers can enter the pretraining stage, so the 0.950 AUC does not necessarily measure generalization to unseen centers. The prospective study (PC1-PC3, collected after development) remains genuinely independent and supports the paradigm, and the internal testing is also informative; therefore the score is moderate rather than severe.

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

The central claim rests on the validity of cytology and histology labels, the representativeness of the 48-center dataset, the transferability of DINOv2 pretext tasks, and the safety of test-time adaptation. The main free parameters are the bag size, contrastive temperature, and adaptation schedule, none of which are fully specified.

free parameters (4)
  • top-k = 50
    Number of top-confidence cell images per class used to construct the WSI bag; chosen in Extended Data Table 16.
  • contrastive temperature tau = not stated
    Temperature in the feature alignment loss (Eq. 2); tunable and not reported.
  • test-time adaptation steps = not stated
    Number of adaptation updates on the test batch is not specified; likely tuned for evaluation.
  • pretraining local crop scale/count = 0.02 scale, 8 crops
    Augmentation settings chosen to match cell instance sizes; affect downstream transfer.
assumptions (5)
  • domain assumption TBS cytology labels are a valid reference standard for screening
    The paper uses cytologist-assigned Bethesda System labels as ground truth for all retrospective evaluation (Methods: Data collection and annotation).
  • domain assumption Histology biopsy is a gold standard for diagnosis validation
    Prospective validation uses 738 histological diagnoses as ground truth (Results: Prospective study).
  • domain assumption The 48 collected centers represent the target clinical population
    The generalization claim assumes these centers capture real-world diversity; all are in China, limiting demographic scope.
  • domain assumption DINOv2 self-supervised features transfer to cytology
    The pretraining stage relies on DINOv2's masked image modeling and alignment losses transfering to cytology patches (Methods: Large-scale self-supervised cytology pretraining).
  • domain assumption Unlabeled test-time adaptation preserves or improves accuracy
    The adaptation stage assumes that aligning test features to source prototypes does not degrade performance under distribution shift (Methods: Test-time adaptation).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation." pith.science (2026). https://pith.science/paper/SYKQZ5ZJ

@misc{pith2026250209662,
  author       = {Pith},
  title        = {Pith review of: Generalizable Cervical Cancer Screening via Large-scale Pretraining and Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYKQZ5ZJ}},
  note         = {Machine review of arXiv:2502.09662}
}
read the original abstract

Cervical cancer is a leading malignancy in female reproductive system. While AI-assisted cytology offers a cost-effective and non-invasive screening solution, current systems struggle with generalizability in complex clinical scenarios. To address this issue, we introduced Smart-CCS, a generalizable Cervical Cancer Screening paradigm based on pretraining and adaptation to create robust and generalizable screening systems. To develop and validate Smart-CCS, we first curated a large-scale, multi-center dataset named CCS-127K, which comprises a total of 127,471 cervical cytology whole-slide images collected from 48 medical centers. By leveraging large-scale self-supervised pretraining, our CCS models are equipped with strong generalization capability, potentially generalizing across diverse scenarios. Then, we incorporated test-time adaptation to specifically optimize the trained CCS model for complex clinical settings, which adapts and refines predictions, improving real-world applicability. We conducted large-scale system evaluation among various cohorts. In retrospective cohorts, Smart-CCS achieved an overall area under the curve (AUC) value of 0.965 and sensitivity of 0.913 for cancer screening on 11 internal test datasets. In external testing, system performance maintained high at 0.950 AUC across 6 independent test datasets. In prospective cohorts, our Smart-CCS achieved AUCs of 0.947, 0.924, and 0.986 in three prospective centers, respectively. Moreover, the system demonstrated superior sensitivity in diagnosing cervical cancer, confirming the accuracy of our cancer screening results by using histology findings for validation. Interpretability analysis with cell and slide predictions further indicated that the system's decision-making aligns with clinical practice. Smart-CCS represents a significant advancement in cancer screening across diverse clinical contexts.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 51 canonical work pages

  1. [1]

    Cancer statistics,

    Rebecca L Siegel, Angela N Giaquinto, and Ahmedin Jemal. Cancer statistics,

    work page

  2. [2]

    Cancer statistics, 2023

    Rebecca L Siegel, Kimberly D Miller, Nikita Sandeep Wagle, and Ahmedin Jemal. Cancer statistics, 2023. CA: a cancer journal for clinicians , 73(1), 2023

    work page 2023

  3. [3]

    Cervical cancer

    Paul A Cohen, Anjua Jhingran, Ana Oaknin, and Lynette Denny. Cervical cancer. The Lancet, 393(10167):169–182, 2019

    work page 2019

  4. [4]

    Cervical cancer kills 300,000 people a year—here’s how to speed up its elimination.Nature, 626(7997):30–32, 2024

    Lynette Denny, Ishu Kataria, Lisa Huang, and Kathleen M Schmeler. Cervical cancer kills 300,000 people a year—here’s how to speed up its elimination.Nature, 626(7997):30–32, 2024

    work page 2024

  5. [5]

    Global strategy to accelerate the elimination of cervical cancer as a public health problem

    World Health Organization et al. Global strategy to accelerate the elimination of cervical cancer as a public health problem . World Health Organization, 2020

    work page 2020

  6. [6]

    Cervical cancer prognosis and survival rates

    National Cancer Institute. Cervical cancer prognosis and survival rates

    work page

  7. [7]

    “smart” cytology: The next generation cytology for precision diagnosis

    Teresa Kim and Jianyu Rao. “smart” cytology: The next generation cytology for precision diagnosis. In Seminars in Diagnostic Pathology , volume 40, pages 95–99. Elsevier, 2023

    work page 2023

  8. [8]

    Discrepancy between cytology and histology in cervical cancer screening: a mul- ticenter retrospective study (kgog 1040)

    Yung-Taek Ouh, Ji Jeong Park, Minjoo Kang, Miseon Kim, Jae Yun Song, So Jin Shin, Seung-Hyuk Shim, Heon Jong Yoo, Maria Lee, Sung-Jong Lee, et al. Discrepancy between cytology and histology in cervical cancer screening: a mul- ticenter retrospective study (kgog 1040). Journal of Korean Medical Science , 36(24), 2021. 23

    work page 2021

Show all 62 references
  1. [9]

    The Bethesda system for reporting cervical cytology: definitions, criteria, and explanatory notes

    Ritu Nayar and David C Wilbur. The Bethesda system for reporting cervical cytology: definitions, criteria, and explanatory notes . Springer, 2015

    2015

  2. [10]

    Sipakmed: A new dataset for feature and image based classification of normal and pathological cervical cells in pap smear images

    Marina E Plissiti, Panagiotis Dimitrakopoulos, Giorgos Sfikas, Christophoros Nikou, O Krikoni, and Antonia Charchanti. Sipakmed: A new dataset for feature and image based classification of normal and pathological cervical cells in pap smear images. In 2018 25th IEEE internatio...

    2018

  3. [11]

    Dual-path network with synergistic grouping loss and evidence driven risk stratification for whole slide cervical image analysis

    Huangjing Lin, Hao Chen, Xi Wang, Qiong Wang, Liansheng Wang, and Pheng- Ann Heng. Dual-path network with synergistic grouping loss and evidence driven risk stratification for whole slide cervical image analysis. Medical Image Analysis, 69:101955, 2021

    2021

  4. [12]

    Deep learning for computational cytology: A survey

    Hao Jiang, Yanning Zhou, Yi Lin, Ronald CK Chan, Jiang Liu, and Hao Chen. Deep learning for computational cytology: A survey. Medical Image Analysis , 84:102691, 2023

    2023

  5. [13]

    Pap-smear benchmark data for pattern classification

    Jan Jantzen, Jonas Norup, Georgios Dounias, and Beth Bjerregaard. Pap-smear benchmark data for pattern classification. Nature inspired smart information systems (NiSIS 2005) , pages 1–9, 2005

    2005

  6. [14]

    A new cervical cytology dataset for nucleus detection and image classification (cervix93) and methods for cervical nucleus detection

    Hady Ahmady Phoulady and Peter R Mouton. A new cervical cytology dataset for nucleus detection and image classification (cervix93) and methods for cervical nucleus detection. arXiv preprint arXiv:1811.09651 , 2018

    2018 arXiv

  7. [15]

    An improved joint optimiza- tion of multiple level set functions for the segmentation of overlapping cervical cells

    Zhi Lu, Gustavo Carneiro, and Andrew P Bradley. An improved joint optimiza- tion of multiple level set functions for the segmentation of overlapping cervical cells. IEEE transactions on image processing , 24(4):1261–1272, 2015

    2015

  8. [16]

    Donet: Deep de-overlapping network for cytology instance segmentation

    Hao Jiang, Rushan Zhang, Yanning Zhou, Yumeng Wang, and Hao Chen. Donet: Deep de-overlapping network for cytology instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 15641–15650, 2023

    2023

  9. [17]

    Gains: Gradient anomaly-aware biomedical instance segmenta- tion

    Runsheng Liu, Hao Jiang, Yanning Zhou, Huangjing Lin, Liansheng Wang, and Hao Chen. Gains: Gradient anomaly-aware biomedical instance segmenta- tion. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 2217–2223. IEEE, 2024

    2024

  10. [18]

    Prediction of tumor origin in cancers of unknown primary origin with cytology-based deep learning

    Fei Tian, Dong Liu, Na Wei, Qianqian Fu, Lin Sun, Wei Liu, Xiaolong Sui, Kathryn Tian, Genevieve Nemeth, Jingyu Feng, et al. Prediction of tumor origin in cancers of unknown primary origin with cytology-based deep learning. Nature Medicine, pages 1–11, 2024

    2024

  11. [19]

    A systematic review of deep learning-based cervical cytol- ogy screening: from cell identification to whole slide image analysis

    Peng Jiang, Xuekong Li, Hui Shen, Yuqi Chen, Lang Wang, Hua Chen, Jing Feng, and Juan Liu. A systematic review of deep learning-based cervical cytol- ogy screening: from cell identification to whole slide image analysis. Artificial Intelligence Review, 56(Suppl 2):2687–2758, 2023

    2023

  12. [20]

    Robust whole slide image analysis for cervical cancer screening using deep learning.Nature communications, 12(1):5639, 2021

    Shenghua Cheng, Sibo Liu, Jingya Yu, Gong Rao, Yuwei Xiao, Wei Han, Wenjie Zhu, Xiaohua Lv, Ning Li, Jing Cai, et al. Robust whole slide image analysis for cervical cancer screening using deep learning.Nature communications, 12(1):5639, 2021

    2021

  13. [21]

    Artificial intelligence enables precision 24 diagnosis of cervical cytology grades and cervical cancer.Nature Communications, 15(1):4369, 2024

    Jue Wang, Yunfang Yu, Yujie Tan, Huan Wan, Nafen Zheng, Zifan He, Luhui Mao, Wei Ren, Kai Chen, Zhen Lin, et al. Artificial intelligence enables precision 24 diagnosis of cervical cytology grades and cervical cancer.Nature Communications, 15(1):4369, 2024

    2024

  14. [22]

    Hybrid ai-assistive diagnostic model permits rapid tbs classification of cervical liquid-based thin-layer cell smears

    Xiaohui Zhu, Xiaoming Li, Kokhaur Ong, Wenli Zhang, Wencai Li, Longjie Li, David Young, Yongjian Su, Bin Shang, Linggan Peng, et al. Hybrid ai-assistive diagnostic model permits rapid tbs classification of cervical liquid-based thin-layer cell smears. Nature communications, 12...

    2021

  15. [23]

    Ai-driven cervical cancer cytolog- ical diagnosis solution based on large scale data collections and annotations: A multi-centre clinical validation

    Weimiao Yu, Fan Zhang, ONG KokHaur, Xinmi Huo, Longjie Li, Peiyao Li, Qihui Wu, Keda Yang, Haoda Lu, Lixiang Wu, et al. Ai-driven cervical cancer cytolog- ical diagnosis solution based on large scale data collections and annotations: A multi-centre clinical validation. 2023

    2023

  16. [24]

    Towards a general-purpose foundation model for computational pathology

    Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850–862, 2024

    2024

  17. [25]

    Towards a gen- eralizable pathology foundation model via unified knowledge distillation

    Jiabo Ma, Zhengrui Guo, Fengtao Zhou, Yihui Wang, Yingxue Xu, Yu Cai, Zhengjie Zhu, Cheng Jin, Yi Lin Xinrui Jiang, Anjia Han, et al. Towards a gen- eralizable pathology foundation model via unified knowledge distillation. arXiv preprint arXiv:2407.18449, 2024

    2024 arXiv

  18. [26]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth´ ee Darcet, Th´ eo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research

  19. [27]

    Holis- tic and historical instance comparison for cervical cell detection

    Hao Jiang, Runsheng Liu, Yanning Zhou, Huangjing Lin, and Hao Chen. Holis- tic and historical instance comparison for cervical cell detection. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , pages 3341–3346. IEEE, 2024

    2024

  20. [28]

    An empirical study of training self- supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self- supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9640–9649, 2021

    2021

  21. [29]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015

    2015

  22. [30]

    An efficient cervical whole slide image analysis framework based on multi-scale semantic and location deep features

    Ziquan Wei, Shenghua Cheng, Junbo Hu, Li Chen, Shaoqun Zeng, and Xiuli Liu. An efficient cervical whole slide image analysis framework based on multi-scale semantic and location deep features. arXiv preprint arXiv:2106.15113 , 2021

    2021 arXiv

  23. [31]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213–229. Springer, 2020

    2020

  24. [32]

    Deformable detr: Deformable transformers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  25. [33]

    atypical

    Christopher J VandenBussche and Zubair W Baloch. The cytologic diagnosis of “atypical”: criteria and controversies. Diagnostic cytopathology, 50(4):143–145, 2022. 25

    2022

  26. [34]

    Attention-based deep mul- tiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep mul- tiple instance learning. In International conference on machine learning , pages 2127–2136. PMLR, 2018

    2018

  27. [35]

    Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learn- ing

    Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learn- ing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2021

    2021

  28. [36]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in neural information processing systems , 34:2136–2147, 2021

    2021

  29. [37]

    Data-efficient and weakly supervised computational pathology on whole-slide images

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Bar- bieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering , 5(6):555–570, 2021

    2021

  30. [38]

    Structured state space models for multiple instance learning in digital pathology

    Leo Fillioux, Joseph Boyd, Maria Vakalopoulou, Paul-Henry Courn` ede, and Ster- gios Christodoulidis. Structured state space models for multiple instance learning in digital pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention , ...

    2023

  31. [39]

    Predicting tumour origin with cytology-based deep learning: hype or hope? Nature Reviews Clinical Oncology, pages 1–2, 2024

    Elie Rassy and Nicholas Pavlidis. Predicting tumour origin with cytology-based deep learning: hype or hope? Nature Reviews Clinical Oncology, pages 1–2, 2024

    2024

  32. [40]

    A new paradigm for cytology- based artificial intelligence-assisted prediction for cancers of unknown primary origins

    Hongrui Li, Shun Wang, Yi Zhang, and Wenbin Li. A new paradigm for cytology- based artificial intelligence-assisted prediction for cancers of unknown primary origins. The Innovation Life , 2(3):100086–1, 2024

    2024

  33. [41]

    Shaoxu Wu, Runnan Shen, Guibin Hong, Yun Luo, Huan Wan, Jiahao Feng, Zeshi Chen, Fan Jiang, Yun Wang, Chengxiao Liao, et al. Development and validation of an artificial intelligence-based model for detecting urothelial carcinoma using urine cytology images: a multicentre, diag...

    2024

  34. [42]

    A foundation model for clinical-grade computational pathology and rare cancers detection

    Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine, ...

    2024

  35. [43]

    A visual-language foundation model for computational pathology

    Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863–874, 2024

    2024

  36. [44]

    The cancer genome atlas pan-cancer analysis project

    John N Weinstein, Eric A Collisson, Gordon B Mills, Kenna R Shaw, Brad A Ozenberger, Kyle Ellrott, Ilya Shmulevich, Chris Sander, and Joshua M Stu- art. The cancer genome atlas pan-cancer analysis project. Nature genetics , 45(10):1113–1120, 2013

    2013

  37. [45]

    Deep learning models for thyroid nodules diagnosis of fine-needle aspiration biopsy: a retrospective, prospective, multicentre study in china

    Jue Wang, Nafen Zheng, Huan Wan, Qinyue Yao, Shijun Jia, Xin Zhang, Sha Fu, Jingliang Ruan, Gui He, Xulin Chen, et al. Deep learning models for thyroid nodules diagnosis of fine-needle aspiration biopsy: a retrospective, prospective, multicentre study in china. The Lancet Digi...

    2024

  38. [46]

    Cytol- ogy assessment can predict survival for patients with metastatic pancreatic neuroendocrine neoplasms

    Carlie S Sigel, Huimin Guo, Keith M Sigel, Ming Zhang, Natasha Rekhtman, Oscar Lin, David S Klimstra, Achim A Jungbluth, and Laura K Tang. Cytol- ogy assessment can predict survival for patients with metastatic pancreatic neuroendocrine neoplasms. Cancer Cytopathology, 125(3):...

    2017

  39. [47]

    Gas- tric cancer with positive peritoneal cytology: survival benefit after induction chemotherapy and conversion to negative peritoneal cytology

    Massimiliano Valletti, Dilmurodjon Eshmuminov, Nicola Gnecco, Chris- tian Alexander Gutschow, Paul Magnus Schneider, and Kuno Lehmann. Gas- tric cancer with positive peritoneal cytology: survival benefit after induction chemotherapy and conversion to negative peritoneal cytolo...

    2021

  40. [48]

    Current role of cytopathology in the molecular and computational era: The perspective of young pathologists

    Alessandro Caputo, Pasquale Pisapia, and Vincenzo L’Imperio. Current role of cytopathology in the molecular and computational era: The perspective of young pathologists. Cancer Cytopathology, 2024

    2024

  41. [49]

    N Paul Ohori, Jacqueline M Cuda, Sheldon I Bastacky, Linwah Yip, Esra Karslioglu-French, Elena M Morariu, Jagdeesh Ullal, Kimberly M Ramonell, Sally E Carty, Yuri E Nikiforov, et al. Molecular-derived risk of malignancy and the related positive call rate of indeterminate thyro...

    2024

  42. [50]

    Openslide: A vendor-neutral software foundation for digital pathology

    Adam Goode, Benjamin Gilbert, Jan Harkes, Drazen Jukic, and Mahadev Satya- narayanan. Openslide: A vendor-neutral software foundation for digital pathology. Journal of pathology informatics , 4(1):27, 2013

    2013

  43. [51]

    A visual–language foundation model for pathology image analysis using medical twitter

    Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual–language foundation model for pathology image analysis using medical twitter. Nature medicine, 29(9):2307–2316, 2023

    2023

  44. [52]

    Transformer-based unsupervised contrastive learning for histopathological image classification

    Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Jun- zhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis , 81:102559, 2022

    2022

  45. [53]

    Less: Label-efficient multi-scale learning for cytological whole slide image screening

    Beidi Zhao, Wenlong Deng, Zi Han Henry Li, Chen Zhou, Zuhua Gao, Gang Wang, and Xiaoxiao Li. Less: Label-efficient multi-scale learning for cytological whole slide image screening. Medical Image Analysis, 94:103109, 2024

    2024

  46. [54]

    Hmil: Hier- archical multi-instance learning for fine-grained whole slide image classification

    Cheng Jin, Luyang Luo, Huangjing Lin, Jun Hou, and Hao Chen. Hmil: Hier- archical multi-instance learning for fine-grained whole slide image classification. IEEE Transactions on Medical Imaging , 2024

    2024

  47. [55]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiao- hua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arX...

    2010 arXiv

  48. [56]

    Yolov3: An incremental improvement

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018

    2018 arXiv

  49. [57]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ ar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision , pages 2980–2988, 2017

    2017

  50. [58]

    Dataset shift in machine learning

    Joaquin Qui˜ nonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning . Mit Press, 2022. 27

    2022

  51. [59]

    Towards launching ai algorithms for cellular pathology into clinical & pharma- ceutical orbits

    Amina Asif, Kashif Rajpoot, David Snead, Fayyaz Minhas, and Nasir Rajpoot. Towards launching ai algorithms for cellular pathology into clinical & pharma- ceutical orbits. arXiv preprint arXiv:2112.09496 , 2021

    2021 arXiv

  52. [60]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7201–7211, 2022

    2022

  53. [61]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gre- gory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems , ...

    2019

  54. [2024]

    CA: a cancer journal for clinicians , 74(1):12–49, 2024

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.