Effortless Vision-Language Model Specialization in Histopathology without Annotation

Jingna Qiu; Jonas Ammeling; Katharina Breininger; Marc Aubreville; Nishanth Jain

REVIEW 4 major objections 5 minor 22 references

Effortless Vision-Language Model Specialization in Histopathology without Annotation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Continued pretraining on keyword-matched image-caption pairs from Quilt1M lets histopathology vision-language models match few-shot performance with no manual labels.

desk verdict Useful empirical study of DAPT/TAPT for histopathology VLMs, but the headline zero-shot gains are missing a leakage control and should not be taken at face value until deduplication is ruled out. read the letter →

arxiv 2508.07835 v1 pith:YOGWXJZC submitted 2025-08-11 cs.CV

Jingna Qiu , Nishanth Jain , Jonas Ammeling , Marc Aubreville , Katharina Breininger This is my paper · ORCID

classification cs.CV

keywords vision-languagemodelshistopathologytaskadaptationcontinuedpretrainingzero-shotclassificationfew-shotlearningimage-captionretrievaldomain-adaptive

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

This paper asks whether a general-purpose histopathology vision-language model can be specialized to a downstream task without a single manual label. The proposed answer is to keep pretraining the model on image-caption pairs pulled from the uncurated Quilt1M database by simple string matching: captions containing the organ name for 'domain-adaptive pretraining' (DAPT), and captions additionally containing the task's class names for 'task-adaptive pretraining' (TAPT). On three classification tasks (breast cancer subtyping on BACH, colorectal polyp typing on MHIST, prostate Gleason grading on SICAP), the paper reports that this label-free continued pretraining substantially raises zero-shot accuracy of both CONCH and QuiltNet, and that with enough unlabeled pairs it reaches the level of the few-shot method CoOp, which needs manually labeled examples. If the claim holds, adapting a pathology VLM to a new task becomes a data-curation chore on text rather than a pathology-expert annotation project.

What carries the argument

The load-bearing machinery is the retrieval-and-continued-pretraining pipeline, in two variants: DAPT (domain-adaptive pretraining) selects image-caption pairs whose captions name the target organ, while TAPT (task-adaptive pretraining) keeps the subset whose captions also contain the task's class names. Selected pairs are ranked by the CONCH model's alignment score - cosine similarity between normalized image and caption embeddings - so the first N pairs are the likely-clean ones; both the image and text encoders are then updated with the dual-encoder contrastive loss from CoCa. The string-matching step is what converts a million-pair, uncurated source into a task-specific, label-free train

What would settle it

Sample about 200 of the TAPT pairs per task and have a pathologist label each as matching or mismatching; the paper reports no such direct measure. If most pairs are mismatches yet continued pretraining still yields the reported gains, the active ingredient is exposure to keyword-filtered statistics, not caption-image fidelity. Complementary control: permute images across captions within the same keyword-filtered set; if accuracy gains persist unchanged, caption-image alignment is not what carries the result.

Watch

Extended reading notes

Core claim

Central claim: keyword-matched domain- and task-relevant image-caption pairs from Quilt1M continue-pretrain a histopathology VLM and substantially improve zero-shot and few-shot performance. DAPT selects captions naming the organ; TAPT narrows to captions containing the class names; pairs are ranked by the CONCH model's alignment score (cosine similarity of image and caption embeddings), then both encoders are updated with a dual-encoder contrastive loss. On QuiltNet, TAPT raises BACH balanced accuracy by 79.62% relative and lifts SICAP weighted kappa from 0.02 to 0.50; CONCH gains more from DAPT, consistent with QuiltNet having already seen Quilt1M in its own pretraining. On MHIST, 16 unlab

Load-bearing premise

The load-bearing premise is that captions whose text contains the organ or class keyword actually describe what is shown in the paired image; the paper never measures that alignment directly, only the downstream accuracy that results from training on these pairs.

Editorial extensions

If this is right

Zero-shot classification improves without any labels on all three tasks, with the largest relative gains where the base model started weakest: QuiltNet's SICAP weighted kappa rises from 0.02 to 0.50 and its BACH balanced accuracy by 79.62% relative.
TAPT can replace few-shot annotation outright on some tasks: on MHIST with QuiltNet, 16 unlabeled pairs matched CoOp with 8 labeled pairs, and 100 unlabeled pairs matched CoOp with 16 labeled pairs.
The adaptation composes with few-shot learning: CoOp applied to the TAPT-adapted model beats CoOp on the original model at shot counts up to 8 (MHIST/QuiltNet).
Which variant helps depends on the base model's history: QuiltNet, already pretrained on Quilt1M, benefits mostly from TAPT, while CONCH, never trained on Quilt1M, benefits more from DAPT - so domain exposure and task exposure are complementary handles.
Since the recipe only needs organ keywords, class keywords, and an image-caption source, it transfers to new histopathology tasks without curation or labeling (the paper's 'task-agnostic' claim).

Reading between the lines

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

The paper never measures whether retrieved captions actually describe their paired images - only downstream accuracy. A pathologist rating a sample of TAPT pairs would directly test whether the gains come from caption-image fidelity or from mere exposure to keyword-filtered statistics.
The paper's own Fig. 2 shows zero-shot gains can decline as lower-ranked pairs are added (SICAP), and its Section 6 proposes thresholding CONCH alignment scores as future work; an explicit rank cutoff or quality gate is a cheap, untested refinement that should concentrate the signal.
If mere exposure is the mechanism, the recipe should transfer to other pathology subdomains with caption sources - cytology, dermatopathology, nephropathology - where class labels are scarce and expert annotation is the bottleneck.
The paper's data concentrate the gains on weak zero-shot baselines (SICAP); on tasks where the base VLM already scores high zero-shot, the label-free margin over few-shot methods may narrow, which the current experiments do not test.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

This is a practical, clearly-written paper. It transplants the DAPT/TAPT idea from RoBERTa to histopathology VLMs, uses a sensible keyword-plus-alignment retrieval recipe over the cleaned Quilt1M subset, and reports consistent gains for two models across three tasks. The code release is a real plus. The observation that QuiltNet, already pretrained on Quilt1M, benefits more from task-adaptive than domain-adaptive pairs, while CONCH benefits more from the latter, is a genuinely interesting and plausible finding.

The soft spots are real but mostly fixable. The stress-test concern is the one I take seriously: there is no deduplication of retrieved pairs against the BACH, MHIST, and SICAP test sets. Quilt1M aggregates PubMed, LAION, and web sources, and all three benchmarks are old enough that their images could appear there. With the reported BACH jump of ~80% and SICAP kappa going from 0.02 to 0.50, the possibility that part of the gain is memorization rather than adaptation needs to be addressed. The authors should add a leakage check—exact and near-duplicate search for test images in the retrieved set, or a held-out source control. Without that, the central equivalence claim to few-shot learning rests on shaky ground.

Other weaknesses are secondary. There is no random-pair baseline, no supervised upper bound, hyperparameter values per training size are not reported, and there are no significance tests beyond min-max shading. The claim that TAPT matches CoOp is based on one task/model configuration. The use of exact class names in both retrieval and zero-shot prompts deserves more discussion, since the vocabulary alignment can inflate gains. Minor editorial issue: Figure 1 contains leftover text ('MICCAI25 next steps') from another venue; that should be cleaned before submission.

Who this is for: anyone working on label-free adaptation of medical VLMs, or on evaluating DAPT/TAPT in new modalities. It deserves a serious referee as an empirical contribution, but the revision should include deduplication, a random-pair control, and reported hyperparameters. I would not cite the headline results until the leakage concern is cleared, but I would engage with the paper and read the revised version carefully.

Referee Report

4 major / 5 minor

Summary. The paper proposes annotation-free specialization of histopathology vision-language models by continued pretraining on retrieved image-caption pairs from Quilt1M. Domain-adaptive (DAPT) pairs are selected by organ keywords, and task-adaptive (TAPT) pairs are a subset filtered by class keywords (Table 1). Retrieved pairs are ranked by CONCH alignment score and used to fine-tune CONCH and QuiltNet with a contrastive loss (Eq. 2). The authors report zero-shot gains on three tasks (BACH, MHIST, SICAP) and show that TAPT can rival or complement CoOp few-shot prompting while requiring no manual labels. Code is released.

Significance. If the results hold, the method is practically valuable: it offers a label-free, task-specific adaptation route for histopathology VLMs, with released code, two foundation models, and three downstream tasks. The use of the Aubreville-cleaned Quilt1M subset and the explicit ranking by alignment score are sensible design choices. However, the central claim of learning from uncurated external data currently lacks controls for data leakage, random-pair baselines, supervised upper bounds, and statistical reliability. These gaps must be addressed before the reported gains can be attributed to the proposed adaptation mechanism rather than to memorization, confounded hyperparameters, or noise.

major comments (4)

[Sec. 4.1, Sec. 3.1, Eq. (2)] No deduplication is reported between the retrieved Quilt1M pairs and the BACH/MHIST/SICAP test sets. Quilt1M aggregates YouTube, PubMed, LAION, and OpenPath content, and the three benchmarks are public; near-duplicate test tiles could therefore appear among the top-ranked retrieved pairs. Since continued pretraining uses full parameter updates with a contrastive loss, even a small number of near-duplicates could inflate the reported gains (e.g., QuiltNet BACH +79.6%, SICAP kappa 0.02→0.50 in Sec. 5.1). The CONCH alignment score used for ranking does not detect test-set overlap. Please add an overlap analysis (e.g., exact/near-duplicate image hashing or feature-space nearest-neighbor search between retrieved pairs and test images) and, ideally, a random-pair or held-out-source control.
[Sec. 5.1, Fig. 2] The only comparison is against the original unadapted model. Without a random-pair or mismatched-caption condition, the improvements cannot be attributed to task relevance rather than to any additional contrastive training or to high alignment scores, which may simply select easier or cleaner pairs. Relatedly, Sec. 5.2's claim that TAPT 'matches the performance of few-shot methods' is not calibrated against a supervised upper bound (e.g., full fine-tuning or CoOp trained on larger labeled sets), so readers cannot judge how much of the few-shot gap remains. Please add random-pair DAPT/TAPT controls and a supervised reference.
[Sec. 4.4] The initial learning rate and weight decay are reported as 'tuned for each training size based on the minimal training loss after 5 epochs,' but the selected values are not given. This is a reproducibility barrier and also a confound: if hyperparameters differ across training sizes, the learning curves in Fig. 2 may reflect hyperparameter choice rather than data quantity. Please report the tuned values for each training size or use a fixed schedule and show sensitivity.
[Fig. 2, Fig. 3] Error visualization is limited to median with min/max over 5 repetitions (Fig. 2) and median over 10 few-shot sets (Fig. 3). Several central comparisons are numerically small (e.g., TAPT vs. CoOp at 8 vs. 16 shots in Fig. 3 left, and DAPT vs. TAPT differences on MHIST). Without confidence intervals or paired significance tests, claims such as 'TAPT significantly enhances CoOp' (Sec. 5.2) and specific improvement percentages are not statistically supported. Please add error bars with confidence intervals and/or paired tests for the main comparisons.

minor comments (5)

[Abstract, Sec. 1, Table 1] The paper calls the method 'task-agnostic,' but TAPT requires task-specific class keywords (Table 1). The pretraining objective is task-agnostic, but the data selection is task-specific. Please qualify the claim accordingly.
[Sec. 3.1] The statement that string matching 'showed that it leads to higher-quality retrievals than similarity searches based on embeddings from PathologyBERT or caption classification using Gemma3' refers to 'preliminary experiments' that are not described. Please provide details or a reference to the supplementary material.
[Sec. 4.4, Fig. 3] Using 'shots' for the number of unlabeled image-caption pairs (N = shots × num classes) is a nonstandard extension of the few-shot convention. It is defined, but the notation should be clearly flagged as 'unlabeled pairs per class' to avoid confusion with labeled examples.
[Sec. 5.2] Few-shot experiments are only shown for MHIST with QuiltNet. The rationale for not presenting CoOp results on other tasks/models should be stated explicitly, since the reader cannot tell whether the pattern is general.
[Figures] Figure 2 uses shaded min/max bands, which are easily confused with confidence intervals; please use standard error plots or explicitly label the bands.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the adaptation claim is an empirical result evaluated on independent labeled benchmarks.

full rationale

The paper does not derive its central claim from its own inputs by construction. TAPT/DAPT pairs are retrieved by string matching on organ/class keywords and ranked by CONCH alignment (Eq. 1), then used to continue pretraining via a contrastive loss (Eq. 2). The evaluation is on held-out labeled test sets (BACH, MHIST, SICAP) using balanced accuracy / kappa, which are not the same as the retrieval scores or the keyword filters. No parameter is fitted to the test labels and then renamed as a prediction. The only self-citation is Aubreville et al. (2024) for the cleaned Quilt1M subset; this is a dataset preprocessing resource, not a justification of the method's effectiveness, and it is externally checkable, so it does not create circularity. The shared vocabulary between TAPT keywords and CONCH zero-shot prompt templates is the intended weak-supervision mechanism, not a definitional equivalence. The paper also acknowledges in Future Work that ensembling alignment measurements may improve data selection, which is a limitation of using a single ranker, not a circular step. Potential test-set leakage from Quilt1M is a data-contamination risk, but it is not a logical reduction of the claim to its inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The experimental findings rest on the assumption that text-matched captions from a noisy database (Quilt1M) correlate with visual content, that ranking pairs by CONCH alignment selects useful data, and that continued contrastive pretraining transfers to downstream classification. There are no derived constants; the main free parameters are the per-task learning rate and weight decay tuned to minimize training loss, plus the chosen prompt context length and LoRA ranks in supplementary experiments.

free parameters (2)

initial learning rate and weight decay per training size
Tuned per training size based on minimal training loss after 5 epochs (Section 4.4); no values are reported, so results may be sensitive to this selection.
CoOp prompt context length M and context type (unified vs CSC) = M=4 or M=16; unified or CSC
Selected as the best combination per task based on evaluation on MHIST (Section 5.2); a design choice rather than a fitted constant.

assumptions (4)

domain assumption String matching on organ names and class names in captions retrieves image-caption pairs relevant to the downstream task.
Section 3.1 selects pairs solely by caption text; Table 1 lists keywords. The validity is only indirectly supported by downstream accuracy.
domain assumption The contrastive objective used in continued pretraining (Eq. 2) improves the alignment of downstream class-relevant features.
The paper updates both encoders with the CoCa dual-encoder contrastive loss, assuming this transfers to zero-shot classification and CoOp few-shot learning.
ad hoc to paper Ranking pairs by CONCH alignment score selects the most useful training pairs.
Introduced in Section 3.1 as a ranking mechanism; preliminary experiments are claimed to support it over PathologyBERT and Gemma3, but no quantitative comparison is given in the main text.
domain assumption The cleaned 232,039-image subset of Quilt1M is sufficiently large and representative to support domain and task adaptation.
Section 4.1 cites Aubreville et al. (2024); the paper relies on that cleaned subset without providing access or checking coverage.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Effortless Vision-Language Model Specialization in Histopathology without Annotation." pith.science (2026). https://pith.science/paper/YOGWXJZC

@misc{pith2026250807835,
  author       = {Pith},
  title        = {Pith review of: Effortless Vision-Language Model Specialization in Histopathology without Annotation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOGWXJZC}},
  note         = {Machine review of arXiv:2508.07835}
}

read the original abstract

Recent advances in Vision-Language Models (VLMs) in histopathology, such as CONCH and QuiltNet, have demonstrated impressive zero-shot classification capabilities across various tasks. However, their general-purpose design may lead to suboptimal performance in specific downstream applications. While supervised fine-tuning methods address this issue, they require manually labeled samples for adaptation. This paper investigates annotation-free adaptation of VLMs through continued pretraining on domain- and task-relevant image-caption pairs extracted from existing databases. Our experiments on two VLMs, CONCH and QuiltNet, across three downstream tasks reveal that these pairs substantially enhance both zero-shot and few-shot performance. Notably, with larger training sizes, continued pretraining matches the performance of few-shot methods while eliminating manual labeling. Its effectiveness, task-agnostic design, and annotation-free workflow make it a promising pathway for adapting VLMs to new histopathology tasks. Code is available at https://github.com/DeepMicroscopy/Annotation-free-VLM-specialization.

Figures

Figures reproduced from arXiv: 2508.07835 by the authors.

**Figure 1.** Retrieved image-caption pairs are utilized in the continued pretraining of a pretrained VLM using a contrastive loss. Their effectiveness is evaluated in both zero-shot and few-shot learning scenarios, in comparison to the original VLM. these models exhibit robust zero-shot performance in histopathology image interpretation, including tile classification and downstream tasks like Whole Slide Image (WSI) segmentati… view at source ↗

**Figure 2.** Comparison of continued pretraining methods, DAPT and TAPT, with the original unadapted model on QuiltNet (top row) and CONCH (bottom row) across three pathology tasks up to 500 shots. The performance of TAPT may terminate earlier if insufficient pairs are retrieved. The results show the median values from five repetitions, with minimum and maximum values shaded. to CONCH, DAPT exhibits higher and more consistent p… view at source ↗

**Figure 3.** Left: Comparison of TAPT and CoOp using unlabeled and labeled training data, respectively. M denotes the prompt’s context length. CSC indicates the use of class-specific context, while the unified context is used otherwise. Right: Comparison of CoOp with and without TAPT across various context types and lengths. Results present the median values from ten randomly selected few-shot sets. the few-shot method CoOp to … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

[1]

Bach: Grand challenge on breast cancer histology images

Guilherme Aresta, Teresa Ara \'u jo, Scotty Kwok, Sai Saketh Chennamsetty, Mohammed Safwan, Varghese Alex, Bahram Marami, Marcel Prastawa, Monica Chan, Michael Donovan, et al. Bach: Grand challenge on breast cancer histology images. Medical image analysis, 56: 0 122--139, 2019

work page 2019
[2]

Model-based cleaning of the quilt-1m pathology dataset for text-conditional image synthesis

Marc Aubreville, Jonathan Ganz, Jonas Ammeling, Christopher C Kaltenecker, and Christof A Bertram. Model-based cleaning of the quilt-1m pathology dataset for text-conditional image synthesis. Medical Imaging with Deep Learning (MIDL), 2024. URL https://openreview.net/forum?id=m7wYKrUjzV

work page 2024
[3]

Clip-adapter: Better vision-language models with feature adapters

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132 0 (2): 0 581--595, 2024

work page 2024
[4]

Histologic grading of prostate cancer: a perspective

Donald F Gleason. Histologic grading of prostate cancer: a perspective. Human pathology, 23 0 (3): 0 273--279, 1992

work page 1992
[5]

Don't stop pretraining: Adapt language models to domains and tasks

Suchin Gururangan, Ana Marasovi \'c , Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964, 2020

arXiv 2004
[6]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

2022
[7]

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 0 (9): 0 2307--2316, 2023

work page 2023
[8]

Quilt-1m: One million image-text pairs for histopathology

Wisdom Ikezogwo, Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. Quilt-1m: One million image-text pairs for histopathology. Advances in neural information processing systems, 36: 0 37995--38017, 2023

work page 2023

Show all 22 references

[9]

Gemma 3 technical report

Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ram \'e , Morgane Rivi \`e re, Louis Rouillard, et al. Gemma 3 technical report. CoRR, 2025

2025
[10]

Roberta: A robustly optimized bert pretraining approach

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019

1907 arXiv
[11]

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 0 (3): 0 863--874, 2024

2024
[12]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

2021
[13]

Pathologybert-pre-trained vs

Thiago Santos, Amara Tariq, Susmita Das, Kavyasree Vayalpati, Geoffrey H Smith, Hari Trivedi, and Imon Banerjee. Pathologybert-pre-trained vs. a new transformer language model for pathology domain. In AMIA annual symposium proceedings, volume 2022, page 962, 2023

2022
[14]

Laion-5b: An open large-scale dataset for training next generation image-text models

Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural informa...

2022
[15]

Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection

Julio Silva-Rodr \' guez, Adri \'a n Colomer, Mar \' a A Sales, Rafael Molina, and Valery Naranjo. Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer methods and programs in b...

2020
[16]

Sus-x: Training-free name-only transfer of vision-language models

Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725--2736, 2023

2023
[17]

A petri dish for histopathology image analysis

Jerry Wei, Arief Suriawinata, Bing Ren, Xiaoying Liu, Mikhail Lisovsky, Louis Vaickus, Charles Brown, Michael Baker, Naofumi Tomita, Lorenzo Torresani, et al. A petri dish for histopathology image analysis. In Artificial Intelligence in Medicine: 19th International Conference ...

2021
[18]

Coca: Contrastive captioners are image-text foundation models

Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022

2022 arXiv
[19]

Low-rank few-shot adaptation of vision-language models

Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1593--1603, 2024

2024
[20]

Tip-adapter: Training-free adaption of clip for few-shot classification

Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European conference on computer vision, pages 493--510. Springer, 2022

2022
[21]

Conditional prompt learning for vision-language models

Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816--16825, 2022 a

2022
[22]

Learning to prompt for vision-language models

Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130 0 (9): 0 2337--2348, 2022 b

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Bach: Grand challenge on breast cancer histology images

Guilherme Aresta, Teresa Ara \'u jo, Scotty Kwok, Sai Saketh Chennamsetty, Mohammed Safwan, Varghese Alex, Bahram Marami, Marcel Prastawa, Monica Chan, Michael Donovan, et al. Bach: Grand challenge on breast cancer histology images. Medical image analysis, 56: 0 122--139, 2019

work page 2019

[2] [2]

Model-based cleaning of the quilt-1m pathology dataset for text-conditional image synthesis

Marc Aubreville, Jonathan Ganz, Jonas Ammeling, Christopher C Kaltenecker, and Christof A Bertram. Model-based cleaning of the quilt-1m pathology dataset for text-conditional image synthesis. Medical Imaging with Deep Learning (MIDL), 2024. URL https://openreview.net/forum?id=m7wYKrUjzV

work page 2024

[3] [3]

Clip-adapter: Better vision-language models with feature adapters

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132 0 (2): 0 581--595, 2024

work page 2024

[4] [4]

Histologic grading of prostate cancer: a perspective

Donald F Gleason. Histologic grading of prostate cancer: a perspective. Human pathology, 23 0 (3): 0 273--279, 1992

work page 1992

[5] [5]

Don't stop pretraining: Adapt language models to domains and tasks

Suchin Gururangan, Ana Marasovi \'c , Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964, 2020

arXiv 2004

[6] [6]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

2022

[7] [7]

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 0 (9): 0 2307--2316, 2023

work page 2023

[8] [8]

Quilt-1m: One million image-text pairs for histopathology

Wisdom Ikezogwo, Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. Quilt-1m: One million image-text pairs for histopathology. Advances in neural information processing systems, 36: 0 37995--38017, 2023

work page 2023