Exploring Foundation Models Fine-Tuning for Cytology Classification

Beno\^it Macq; Isabelle Salmon; Karim El Khoury; Manon Dausort; Maxime Zanella; Tiffanie Godelaine

REVIEW 4 major objections 6 minor 34 references

Exploring Foundation Models Fine-Tuning for Cytology Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning a general-purpose CLIP encoder with LoRA outperforms the previous state of the art on a 25-class cervical cytology benchmark while training 62 times fewer parameters and using 70% of the data.

desk verdict Useful empirical comparison, but the 70%-data SOTA claim needs the HiCervix protocol spelled out before it can be checked. read the letter →

arxiv 2411.14975 v1 pith:EVTE5EK5 submitted 2024-11-22 eess.IV cs.AIcs.CVq-bio.QM

Manon Dausort , Tiffanie Godelaine , Maxime Zanella , Karim El Khoury , Isabelle Salmon , Benoît Macq This is my paper · ORCID

classification eess.IVcs.AIcs.CVq-bio.QM

keywords CytologyclassificationFoundationmodelsLow-rankadaptationLoRAFew-shotlearningParameter-efficientfine-tuningHiCervixCervical

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

The paper tries to establish that existing foundation models—large models pre-trained on broad image data—can be adapted with low-rank updates to reach state-of-the-art cytology classification without training a cytology-specific model from scratch. Across four cytology datasets, the authors find that updating only the classifier head leaves accuracy behind, whereas applying LoRA to the visual backbone consistently improves performance, especially when only a few labeled examples per class are available. On the large 25-class HiCervix benchmark, a CLIP vision encoder with a ViT-L/14 backbone fine-tuned by LoRA is reported to outperform HierSwin, the prior state of the art, using 70% of the dataset and only 3.1 million trainable parameters. The broader point is that general-purpose vision-language pretraining combined with parameter-efficient adaptation may be a practical path for cytology, where annotated data is scarce.

What carries the argument

Low-Rank Adaptation (LoRA) is the central mechanism; it models each weight update as a product of two small matrices $\Delta W = BA$ added to frozen pre-trained weights, so only a small number of parameters are trained and inference cost stays identical to the original model. The paper applies LoRA to the attention matrices of vision transformers; in the few-shot experiments it uses rank 2 on query and value matrices, and for the final HiCervix experiment it applies rank 16 to query, key, value, and output matrices of a ViT-L/14 CLIP encoder. This mechanism is what allows a general-purpose CLIP backbone to be adapted without full fine-tuning, and the paper attributes the gains over head-only linear probing to updating the backbone's representations in low-data regimes.

What would settle it

Train the same CLIP ViT-L/14 LoRA model on exactly the official HiCervix benchmark protocol that HierSwin used and compare top-1 accuracy on the identical test set; if the 70%-data run does not reach HierSwin's reported number, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that fine-tuning a foundation model's backbone with Low-Rank Adaptation (LoRA), rather than only its classification head, is the key to transferring general visual knowledge to cytology. In few-shot regimes from 1 to 50 shots per class, LoRA-tuned CLIP outperforms its own head-only baseline with one or two examples per class on three datasets, and on HiCervix it needs 50 shots (4.4% of the data) to match head-only performance. The strongest result is on HiCervix's third-level 25-class label hierarchy: CLIP's ViT-L/14 encoder with rank-16 LoRA on query, key, value, and output matrices reaches state-of-the-art accuracy when trained on 70% of the dataset, with 3.1 million trainable parameters, roughly 62 times fewer than HierSwin. That is the discovery the paper is trying to establish: a general-purpose encoder, lightly adapted, can beat a cytology-specific hierarchical transformer while using less data and far fewer tunable parameters.

Load-bearing premise

The comparison to HierSwin is valid: the authors' HiCervix split, the 25-class label set, the preprocessing, and the evaluation metric must match the protocol that produced HierSwin's reported accuracy, and the paper does not specify how its 5-100% data subsets were sampled.

Editorial extensions

If this is right

For cytology tasks with very few labeled images, LoRA fine-tuning of a general-purpose vision-language backbone is a stronger choice than freezing the backbone and training a linear head.
A model trained on 70% of HiCervix can match the reported state-of-the-art accuracy, so a meaningful fraction of labeling and annotation effort could be saved.
Training only 3.1 million parameters instead of the full model keeps the adaptation practical, and LoRA adds no extra computation at inference time.
Using all available HiCervix data pushes accuracy to 80.23%, showing that performance continues to improve as more labeled cytology data is added.
Histology-pretrained models work well as frozen feature extractors, but when labeled data is scarce, broader general-purpose pretraining adapts better through LoRA.

Reading between the lines

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

The paper does not test whether LoRA rank and the choice of which attention matrices to adapt matter more as the label hierarchy grows; a systematic sweep could yield even better data efficiency on HiCervix.
If the comparison protocol is confirmed, this suggests that for cytology, specialized histology pretraining is not automatically the best starting point, because cytology images lack the tissue architecture that histology models are built around.
A natural extension the authors leave open is using the text side of CLIP as well, since only the vision encoder is fine-tuned here; adding language supervision might improve the harder 25-class task further.

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

Short version: the paper is a legitimate empirical comparison of five foundation models on four cytology datasets, testing LoRA against a frozen-backbone linear head. The central finding -- LoRA beats head-only tuning in few-shot regimes -- is consistent across all four datasets, and the claim that a general-purpose CLIP ViT-L/14 with LoRA reaches the reported HierSwin accuracy on HiCervix using 70% of the data and 3.1 million trainable parameters is, if true, a useful result for the cytology community. The code and datasets are public, which makes the work reproducible in principle.

What it does well: the evaluation is systematic. Same training protocol, three seeds, validation-based learning-rate selection, and the few-shot curves in Fig. 1 tell a coherent story. The observation that histology-specific models underperform the more general CLIP when LoRA-fine-tuned in low-data settings is non-obvious and worth reporting.

The soft spots are missing details, not conceptual errors. The SOTA comparison with HierSwin rests on an unverified protocol: the paper never states how the HiCervix data was split, whether the percentage subsets were drawn from the whole set or only the official training partition, whether sampling was stratified by class or center, or what preprocessing and metric HierSwin's reported number uses. Fig. 2 plots HierSwin as an unlabeled horizontal line, so the claimed crossover at 70% cannot be checked. Three-seed means without error bars leave the margin unknown. The abstract overstates the result by claiming state-of-the-art on 'simple and complex tasks' when the comparison is only on HiCervix.

The LoRA hyperparameters (rank 2 on query/value, then rank 16 on all attention matrices) come from a co-authored prior paper [24] -- a self-citation, but not a fatal one. The choice is plausible and the central benchmark results are not forced by it. The circularity burden is low: the 70% crossing is read off the data, not derived from a fitted model.

If the protocol details are added -- exact split, sampling, preprocessing, and the numerical HierSwin accuracy -- the headline claim becomes checkable. As written, it is not independently verifiable, which is a fixable but real problem.

Send to peer review. The empirical core deserves referee time; the missing details are standard revision requests. I'd accept with major revisions, conditional on the protocol being fully specified.

Referee Report

4 major / 6 minor

Summary. The paper explores fine-tuning five foundation models (CLIP, QUILT, BiomedCLIP, UNI, ViT) on four public cytology classification datasets (BCFC, MLCC, SIPaKMeD, HiCervix). Three experiments are reported: (1) frozen backbones with a linear classifier, (2) LoRA fine-tuning of query/value matrices in a few-shot setting (1 to 50 shots per class), and (3) scaling LoRA fine-tuning on HiCervix across data proportions from 5% to 100%. The central claims are that LoRA fine-tuning consistently beats linear probing in few-shot regimes and that CLIP with a ViT-L/14 backbone, fine-tuned with LoRA on 70% of HiCervix, achieves state-of-the-art accuracy while training only 3.1 million parameters (62x fewer than HierSwin) and using less data. The code is made publicly available.

Significance. If the claims hold, the paper provides a practical and parameter-efficient recipe for cytology classification: a general-purpose CLIP encoder adapted with LoRA outperforms a cytology-specific Swin Transformer (HierSwin) on a large, complex 25-class dataset while using fewer trainable parameters and less training data. The paper also provides a broad comparison across five foundation models and four datasets, with results averaged over three seeds, which is useful evidence for transfer-learning choices in cytology. The main strength is the empirical breadth and the reproducible code release; the main weakness is that the headline state-of-the-art claim rests on an under-specified comparison protocol that is not currently independently checkable.

major comments (4)

[§4.3, Fig. 2] The headline claim that fine-tuning on 70% of HiCervix achieves state-of-the-art performance is not checkable as written. The paper does not state whether the 5-100% subsets are sampled from the official training partition or from the entire HiCervix dataset, whether the sampling is random, class-stratified, or center-stratified, what the patch preprocessing and input resolution are, and what numerical HierSwin accuracy is used as the baseline. If the percentage subsets include test images or have different class frequencies than the official benchmark, the comparison to HierSwin is invalid. Please report the exact split, sampling scheme, preprocessing, evaluation protocol, and the numerical HierSwin accuracy.
[§4.3, Fig. 2] The crossover between CLIP-LoRA and HierSwin is presented with three-seed means but without error bars or standard deviations, and the HierSwin line is plotted without a numerical value. The margin between the 70% data point and the state-of-the-art line cannot be assessed, so it is unclear whether the claimed superiority is statistically meaningful. Please include the standard deviations or error bars and give the numerical value of the HierSwin reference accuracy in the text or figure caption.
[§4.2, Fig. 1] The few-shot comparison in Experiment 2 omits numerical results and error bars for the 1-50 shots curves, which makes it difficult to verify the claim that LoRA 'consistently' outperforms the linear classifier and to locate the HiCervix crossing at 50 shots. In addition, the LoRA hyperparameters (rank 2, applied only to query and value) are taken from the authors' prior work [24] without an ablation or justification; since these choices partly define the method being compared, the claim would be stronger with an ablation study or at least a sensitivity check.
[§4.3, Experiment 3 details] The LoRA configuration in Experiment 3 changes to rank 16 on query, value, key, and output for the ViT-L/14 backbone, but no hyperparameter selection procedure or training details (learning rate, epochs, optimizer, learning-rate schedule, validation criteria) are described for the data-proportion curves. Since this experiment is the basis of the state-of-the-art claim, the paper should state these details explicitly or clarify that they match the protocol of Experiment 2.

minor comments (6)

[§4.2, last paragraph] The sentence 'fine-tuning the entire backbone consistently yields higher accuracy' is inaccurate because the experiments use LoRA, which updates only low-rank matrices on query and value, not the entire backbone; please rephrase to 'fine-tuning the backbone with LoRA'.
[Table 1] The dataset name is spelled 'SIPaKMed' in Table 1 but 'SIPaKMeD' in the text and elsewhere; please standardize the spelling.
[§3.3, Ref. [7]] The ViT model is cited to Wu et al. (arXiv:2006.03677), which is not the original ViT paper; the standard reference is Dosovitskiy et al., 'An Image is Worth 16x16 Words' (ICLR 2021), or the authors should clarify which ViT variant they actually use.
[§4.1] The statement that 'UNI with a fine-tuned classifier achieves accuracies consistent with those reported in the literature' lacks a citation or a comparison table; please add the reference and the corresponding numbers.
[Fig. 2 caption] The horizontal line for HierSwin has no numerical label; please add the accuracy value in the caption or in the text, since readers cannot determine the reference level from the figure alone.
[Abstract] The abstract claims 'state-of-the-art results on both simple and complex classification tasks', but the state-of-the-art comparison is only made for HiCervix in Experiment 3; please qualify the claim to avoid overstatement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the LoRA and data-percentage results are empirical measurements, and the only self-citation (LoRA rank choice) is not load-bearing.

full rationale

The paper's claimed derivation chain is experimental rather than formal. In Experiment 2, the comparison between LoRA fine-tuning and linear-head fine-tuning is measured on held-out test sets across shot counts; no fitted parameter is reused as a predicted result. In Experiment 3, the claim that CLIP with LoRA reaches HierSwin-level accuracy at 70% of the HiCervix dataset is read directly from the measured accuracy curve in Fig. 2; it is not generated by an equation fit to the same data. The only authors' self-citation, [24], justifies setting the LoRA rank to 2 and applying LoRA to query/value matrices. That is a hyperparameter choice, not a theorem whose conclusion coincides with the paper's headline result, so it does not force the outcome. The main substantive gap is external comparability, not circularity: the paper does not state how HiCervix was split, how the 5-100% subsets were sampled, or the numerical HierSwin accuracy plotted in Fig. 2, so the state-of-the-art claim cannot be independently checked. That is a correctness and verifiability concern, not a circular-derivation concern.

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

The paper contributes an empirical benchmark; its claims rest on many unstated training and evaluation choices rather than on a mathematical derivation. The most consequential free parameter is the post hoc 70% threshold, and the most consequential axiom is split and protocol comparability with HierSwin.

free parameters (5)

LoRA rank (Experiment 2) = 2
Chosen from co-author prior work [24] rather than ablated on cytology data; affects the magnitude of LoRA gains over linear probing.
LoRA rank and applied matrices (Experiment 3) = 16 on query, key, value, output projections
Changed from rank 2 on query and value only; no ablation justifies this switch, and it changes the trainable parameter count (3.1 million).
Learning rates = not reported
Selected per model-shots-dataset on a validation set; values are absent, so the linear versus LoRA comparison cannot be audited for equal tuning effort.
Number of training epochs = not reported
No epoch count or early stopping criterion is given; training budget may differ across configurations.
HiCervix 70% proportion = 70%
Identified post hoc from the accuracy versus percentage curve as the smallest proportion reaching HierSwin-level accuracy; this is a fitted threshold, not a pre-registered claim.

assumptions (4)

standard math LoRA is a faithful parameter-efficient fine-tuning method that can approximate downstream task updates (Hu et al. [17]).
The paper relies on the LoRA formulation (Eq. 2) without re-deriving it.
domain assumption The validation set used for learning-rate selection does not leak into the test set and is representative.
Section 4 states best learning rate is chosen on validation and final accuracy is measured on test, but no split details are provided.
domain assumption HiCervix third-level labels and evaluation protocol coincide with HierSwin [15].
The SOTA comparison in Section 4.3 assumes comparability; the paper does not describe its split or preprocessing, nor confirm it matches Cai et al.
domain assumption Input image preprocessing (patch size, normalization, augmentation) is consistent across models and datasets.
No preprocessing details are given anywhere; if models expect different input statistics, comparisons could be biased.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Foundation Models Fine-Tuning for Cytology Classification." pith.science (2026). https://pith.science/paper/EVTE5EK5

@misc{pith2026241114975,
  author       = {Pith},
  title        = {Pith review of: Exploring Foundation Models Fine-Tuning for Cytology Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EVTE5EK5}},
  note         = {Machine review of arXiv:2411.14975}
}

read the original abstract

Cytology slides are essential tools in diagnosing and staging cancer, but their analysis is time-consuming and costly. Foundation models have shown great potential to assist in these tasks. In this paper, we explore how existing foundation models can be applied to cytological classification. More particularly, we focus on low-rank adaptation, a parameter-efficient fine-tuning method suited to few-shot learning. We evaluated five foundation models across four cytological classification datasets. Our results demonstrate that fine-tuning the pre-trained backbones with LoRA significantly improves model performance compared to fine-tuning only the classifier head, achieving state-of-the-art results on both simple and complex classification tasks while requiring fewer data samples.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

[24]

Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models,

Elad Ben Zaken, Shauli Ravfogel, et al., “Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models,” arXiv preprint arXiv:2106.10199, 2021

arXiv 2021
[1]

INTRODUCTION Cytology slides are vital for diagnosing and staging cancer, offering detailed views of abnormal cells to guide treatment decisions [1]. Analyzing these slides is labor- intensive and costly, leading to delays in the reporting process, making au- tomation essential to improve both classification efficiency and accuracy [2]. An effective appro...

work page Pith review arXiv 2024
[2]

RELATED WORK 2.1. Cytology classification Many cytology classification studies use pre-trained CNN models as feature extractors, typically fine-tuning only the classification layer added on top of the backbone [19, 20]. For example, UC ¸ Aet al. [21] compared various CNN ar- chitectures, finding ResNet50 effective for body cavity cy- tologies. Yaman and T...

work page
[3]

Fine-tuning methods To adapt FMs to cytology classification, we compare two fine-tuning strategies

EXPERIMENTAL SETUP 3.1. Fine-tuning methods To adapt FMs to cytology classification, we compare two fine-tuning strategies. Linear classifier. A common fine-tuning approach is to lever- age a pre-trained model θ to extract key features z = fθ(x) from an input x and to add a linear classifier on top. Dur- ing training, only the weights of the linear layer ...

work page 2012
[4]

EXPERIMENTS Results are averaged over 3 seeds. Top-1 accuracy is evalu- ated on the validation set to determine the best learning rate for each model-shots-dataset combination, and final reported performances are measured on the test set of each dataset. The batch size is set to 32. Table 1. Mean accuracy of fine-tuned classifiers, evaluated on five model...

work page
[5]

For instance, 2 shots of the SIPaKMeD dataset results a total of 10 examples (two examples per class)

The model is trained in a few-shot setting with shots per class ranging from 1, 2, 4, 8, 16, to 50. For instance, 2 shots of the SIPaKMeD dataset results a total of 10 examples (two examples per class). Results. The results can be seen in Fig.1. CLIP fine-tuned with LoRA outperforms the linear classifier with only one or two samples per class on the first...

work page
[6]

CONCLUSION This study investigates the potential of foundation models for cytology. We show that fine-tuning with LoRA significantly improves performance over classifier-only fine-tuning, with particularly strong results in few-shot settings where labeled data is limited, and further gains when more data is available. Our findings suggest that while histo...

work page
[7]

COMPLIANCE WITH ETHICAL STANDARDS This is a study for which no ethical approval was required

work page

Show all 34 references

[8]

Dausort and T

ACKNOWLEDGMENTS M. Dausort and T. Godelaine are funded by the MedReSyst project, supported by FEDER and the Walloon Region. M. Zanella is funded by the Walloon region under grant No. 2010235 (ARIAC by DIGITALW ALLONIA4.AI). Computa- tional resources were made available on the ...
[9]

The history of cervical screening i: the pap. test,

Patricia A Shaw, “The history of cervical screening i: the pap. test,” Journal SOGC, vol. 22, no. 2, pp. 110–114, 2000

2000
[11]

Diversity in machine learning,

Zhiqiang Gong, Ping Zhong, et al., “Diversity in machine learning,” Ieee Access, vol. 7, pp. 64323–64350, 2019

2019
[12]

Deeppap: deep convolutional net- works for cervical cell classification,

Ling Zhang, Le Lu, et al., “Deeppap: deep convolutional net- works for cervical cell classification,”IEEE journal of biomed- ical and health informatics , vol. 21, no. 6, pp. 1633–1643, 2017

2017
[13]

A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities,

Yisheng Song, Ting Wang, et al., “A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities,” ACM Computing Surveys, vol. 55, no. 13s, pp. 1–40, 2023

2023
[14]

Learning transferable visual models from natural language supervision,

Alec Radford, Jong Wook Kim, et al., “Learning transferable visual models from natural language supervision,” in Inter- national conference on machine learning . PMLR, 2021, pp. 8748–8763

2021
[15]

Visual transformers: Token- based image representation and processing for computer vi- sion,

Bichen Wu, Chenfeng Xu, et al., “Visual transformers: Token- based image representation and processing for computer vi- sion,” arXiv preprint arXiv:2006.03677, 2020

2006 arXiv
[16]

Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs,

Sheng Zhang, Yanbo Xu, et al., “Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs,” arXiv preprint arXiv:2303.00915, 2023

2023 arXiv
[17]

A visual–language foun- dation model for pathology image analysis using medical twit- ter,

Zhi Huang, Federico Bianchi, et al., “A visual–language foun- dation model for pathology image analysis using medical twit- ter,” Nature medicine, vol. 29, no. 9, pp. 2307–2316, 2023

2023
[18]

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

Wisdom Ikezogwo, Saygin Seyfioglu, et al., “Quilt-1m: One million image-text pairs for histopathology,” Advances in neu- ral information processing systems, vol. 36, 2024

2024
[19]

A visual-language foundation model for computational pathology,

Ming Y Lu, Bowen Chen, et al., “A visual-language foundation model for computational pathology,”Nature Medicine, vol. 30, no. 3, pp. 863–874, 2024

2024
[20]

Towards a general- purpose foundation model for computational pathology,

Richard J Chen, Tong Ding, et al., “Towards a general- purpose foundation model for computational pathology,” Na- ture Medicine, vol. 30, no. 3, pp. 850–862, 2024

2024
[21]

Boosting vision- language models for histopathology classification: Predict all at once,

Maxime Zanella, Fereshteh Shakeri, et al., “Boosting vision- language models for histopathology classification: Predict all at once,” in International Workshop on Foundation Models for General Medical AI. Springer, 2024, pp. 153–162

2024
[22]

Improving mitosis detection on histopathology images using large vision-language mod- els,

Ruiwen Ding, James Hall, et al., “Improving mitosis detection on histopathology images using large vision-language mod- els,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

2024
[23]

Hicervix: An extensive hierarchi- cal dataset and benchmark for cervical cytology classification,

De Cai, Jie Chen, et al., “Hicervix: An extensive hierarchi- cal dataset and benchmark for cervical cytology classification,” IEEE Transactions on Medical Imaging, 2024

2024
[25]

Lora: Low-rank adaptation of large language models,

Edward J Hu, Yelong Shen, et al., “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

2021 arXiv
[26]

Parameter-efficient fine- tuning for medical image analysis: The missed opportunity,

Raman Dutt, Linus Ericsson, et al., “Parameter-efficient fine- tuning for medical image analysis: The missed opportunity,” arXiv preprint arXiv:2305.08252, 2023

2023 arXiv
[27]

Deep learning for computa- tional cytology: A survey,

Hao Jiang, Yanning Zhou, et al., “Deep learning for computa- tional cytology: A survey,” Medical Image Analysis, vol. 84, pp. 102691, 2023

2023
[28]

A systematic review of deep learning-based cervical cytology screening: from cell identifi- cation to whole slide image analysis,

Peng Jiang, Xuekong Li, et al., “A systematic review of deep learning-based cervical cytology screening: from cell identifi- cation to whole slide image analysis,” Artificial Intelligence Review, vol. 56, no. Suppl 2, pp. 2687–2758, 2023

2023
[29]

Comparison of deep learning models for body cavity fluid cytology images classification,

Murat UC ¸ A, Buket Kaya, et al., “Comparison of deep learning models for body cavity fluid cytology images classification,” in 2022 International Conference on data analytics for business and industry (ICDABI). IEEE, 2022, pp. 151–155

2022
[30]

Exemplar pyramid deep feature extraction based cervical cancer image classification model using pap-smear images,

Orhan Yaman and Turker Tuncer, “Exemplar pyramid deep feature extraction based cervical cancer image classification model using pap-smear images,” Biomedical Signal Process- ing and Control, vol. 73, pp. 103428, 2022

2022
[31]

Parameter-efficient fine-tuning for large models: A comprehensive survey,

Z. Han, C. Gao, J. Liu, et al., “Parameter-efficient fine-tuning for large models: A comprehensive survey,” 2024

2024
[32]

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 Recog- nition, 2024, pp. 1593–1603

2024
[33]

Body cavity fluid cytology images,

P. Sanyal, “Body cavity fluid cytology images,” https://www.kaggle.com/datasets/cmacus/ body-cavity-fluid-cytology-images
[34]

Liquid based-cytology pap smear dataset for automated multi-class diagnosis of pre- cancerous and cervical cancer lesions,

Elima Hussain, Lipi B Mahanta, et al., “Liquid based-cytology pap smear dataset for automated multi-class diagnosis of pre- cancerous and cervical cancer lesions,” Data in brief, vol. 30, pp. 105589, 2020

2020
[35]

Sipakmed: A new dataset for feature and image based classi- fication of normal and pathological cervical cells in pap smear images,

Marina E Plissiti, Panagiotis Dimitrakopoulos, et al., “Sipakmed: A new dataset for feature and image based classi- fication of normal and pathological cervical cells in pap smear images,” in 2018 25th IEEE international conference on image processing (ICIP). IEEE, 2018, pp. 3144–3148

2018

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [24]

Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models,

Elad Ben Zaken, Shauli Ravfogel, et al., “Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models,” arXiv preprint arXiv:2106.10199, 2021

arXiv 2021

[2] [1]

INTRODUCTION Cytology slides are vital for diagnosing and staging cancer, offering detailed views of abnormal cells to guide treatment decisions [1]. Analyzing these slides is labor- intensive and costly, leading to delays in the reporting process, making au- tomation essential to improve both classification efficiency and accuracy [2]. An effective appro...

work page Pith review arXiv 2024

[3] [2]

RELATED WORK 2.1. Cytology classification Many cytology classification studies use pre-trained CNN models as feature extractors, typically fine-tuning only the classification layer added on top of the backbone [19, 20]. For example, UC ¸ Aet al. [21] compared various CNN ar- chitectures, finding ResNet50 effective for body cavity cy- tologies. Yaman and T...

work page

[4] [3]

Fine-tuning methods To adapt FMs to cytology classification, we compare two fine-tuning strategies

EXPERIMENTAL SETUP 3.1. Fine-tuning methods To adapt FMs to cytology classification, we compare two fine-tuning strategies. Linear classifier. A common fine-tuning approach is to lever- age a pre-trained model θ to extract key features z = fθ(x) from an input x and to add a linear classifier on top. Dur- ing training, only the weights of the linear layer ...

work page 2012

[5] [4]

EXPERIMENTS Results are averaged over 3 seeds. Top-1 accuracy is evalu- ated on the validation set to determine the best learning rate for each model-shots-dataset combination, and final reported performances are measured on the test set of each dataset. The batch size is set to 32. Table 1. Mean accuracy of fine-tuned classifiers, evaluated on five model...

work page

[6] [5]

For instance, 2 shots of the SIPaKMeD dataset results a total of 10 examples (two examples per class)

The model is trained in a few-shot setting with shots per class ranging from 1, 2, 4, 8, 16, to 50. For instance, 2 shots of the SIPaKMeD dataset results a total of 10 examples (two examples per class). Results. The results can be seen in Fig.1. CLIP fine-tuned with LoRA outperforms the linear classifier with only one or two samples per class on the first...

work page

[7] [6]

CONCLUSION This study investigates the potential of foundation models for cytology. We show that fine-tuning with LoRA significantly improves performance over classifier-only fine-tuning, with particularly strong results in few-shot settings where labeled data is limited, and further gains when more data is available. Our findings suggest that while histo...

work page

[8] [7]

COMPLIANCE WITH ETHICAL STANDARDS This is a study for which no ethical approval was required

work page