{"id":"d86b72b1-3469-4fa5-abeb-f277e5f00c04","arxiv_id":"2411.14975","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LoRA fine-tuning of pretrained vision models beats head-only fine-tuning on four cytology benchmarks, and CLIP with LoRA matches the HiCervix state of the art using 70% of the data.","lead":"The paper tests five pretrained image models on four collections of microscope images of cells (cytology) and finds that lightly fine-tuning the whole model with a technique called LoRA beats only training the final classifier. If the result holds, AI cancer screening for cytology can reach top accuracy with less labeled data and far fewer trainable parameters, lowering the cost of building such tools.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 70%-data SOTA claim depends on an unverified HiCervix evaluation protocol; without official split/sampling details and an explicit HierSwin accuracy number, the headline comparison cannot be checked.","rationale":"I read the paper as an empirical study with real strengths: it releases code, averages across three seeds, evaluates five foundation models on four datasets, and reports results that are internally consistent. The central claim, however, is the Experiment 3 result that a CLIP ViT-L/14 fine-tuned with LoRA on 70% of HiCervix outperforms HierSwin, the prior state of the art, while using fewer parameters and fewer data. For that claim to hold, the comparison must be protocol-identical: same data split, same 25-class label hierarchy, same preprocessing, and same evaluation metric on the same test set. The paper does not specify any of these details for the percentage-based curves, neither how subsets were sampled nor where the HierSwin accuracy value comes from. The horizontal line in Fig. 2 is unlabeled and unquantified in the text, and no error bars are shown. This is not an accusation of bad practice; it is a missing-reporting issue that prevents verification. If the official split is used and subsets are drawn from the training partition with class stratification, the SOTA claim may well be true, but as written the reader cannot rule out the more mundane explanation that the comparison is not apples-to-apples. The reader's weakest assumption identified exactly this protocol-matching issue, and I agree. I therefore recommend keeping the CONDITIONAL verdict unchanged: the concern is substantive and fixable, but it does not by itself warrant rejection without first running the concrete protocol check.","tokens_in":6746,"tokens_out":4642,"duration_ms":47649,"concrete_test":"Re-run Experiment 3 on the official HiCervix benchmark: sample 70% of the official training partition via class-stratified random draws with the same three seeds, train CLIP ViT-L/14 plus LoRA (rank 16 on query, key, value, and output matrices) using the same patch preprocessing as HierSwin, and evaluate on the official test set used by Cai et al. Report the 3-seed mean and standard deviation along with the exact HierSwin accuracy from reference [15]. If the mean does not exceed HierSwin's number by more than one standard deviation, the SOTA claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Experiment 3's state-of-the-art claim rests on an apples-to-apples comparison with HierSwin: the same official train/val/test split, the same third-level 25-class labels, the same patch preprocessing, and the same top-1 metric. None of these are described for the 5-100% curves. In particular, the paper never states whether each 'percentage of the dataset' is a random draw from the entire HiCervix set or from the official training partition, nor whether the sampling is stratified by class or center. If the draw includes test images, the 70% result is inflated; if it is an unstratified random subsample, class frequencies differ from HierSwin's benchmark and top-1 accuracy is no longer comparable. Fig. 2 also plots HierSwin as an unlabeled horizontal line without a numerical value, so the claimed crossover cannot be checked. The 3-seed means are shown without error bars, so even a valid protocol leaves the margin of the crossover unknown. The 'SOTA with 70% data' claim therefore rests on an unverified protocol; this is fixable by reporting the exact split, sampling, preprocessing, and HierSwin accuracy, but as written the central claim is not independently checkable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6961,"tokens_out":4361,"duration_ms":40470,"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":[{"comment":"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.","section":"§4.3, Fig. 2"},{"comment":"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.","section":"§4.3, Fig. 2"},{"comment":"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.","section":"§4.2, Fig. 1"},{"comment":"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.","section":"§4.3, Experiment 3 details"}],"minor_comments":[{"comment":"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'.","section":"§4.2, last paragraph"},{"comment":"The dataset name is spelled 'SIPaKMed' in Table 1 but 'SIPaKMeD' in the text and elsewhere; please standardize the spelling.","section":"Table 1"},{"comment":"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.","section":"§3.3, Ref. [7]"},{"comment":"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.","section":"§4.1"},{"comment":"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.","section":"Fig. 2 caption"},{"comment":"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.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper reads like a workshop-style short paper that is concise but missing key protocol details for its strongest claim. The main concerns are the unverified HiCervix split/sampling and the missing HierSwin reference number; these are fixable with a clear protocol description and a numerical baseline. The self-citation for LoRA hyperparameters is acceptable but should be supplemented by a justification or an ablation. If the authors provide the missing details, the paper would be a useful empirical contribution to PEFT for cytology classification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat 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.\n\nThe 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.\n\nThe 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.\n\nIf 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.\n\nSend 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.","headline":"Useful empirical comparison, but the 70%-data SOTA claim needs the HiCervix protocol spelled out before it can be checked.","tokens_in":7581,"tokens_out":3190,"would_cite":false,"duration_ms":29916,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["Cytology classification","Foundation models","Low-rank adaptation","LoRA","Few-shot learning","Parameter-efficient fine-tuning","HiCervix","Cervical cytology"],"falsifier":"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.","tokens_in":6514,"feed_emoji":"🔬","tokens_out":10114,"duration_ms":81200,"temperature":0.7,"pith_summary":"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.","feed_headline":"LoRA-tuned CLIP tops cervical cytology benchmark with 70% data","feed_subtitle":"General-purpose CLIP with LoRA beats a cytology-specific transformer using fewer parameters and less data.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines LoRA, the low-rank adaptation mechanism that carries all fine-tuning experiments.","marker":"[17]"},{"why":"Provides the rank-2 LoRA recipe for CLIP that the few-shot experiments follow.","marker":"[24]"},{"why":"Supplies the HiCervix dataset, its 25-class hierarchy, and the HierSwin state-of-the-art baseline that Experiment 3 must beat.","marker":"[15]"},{"why":"Provides CLIP, the general-purpose vision-language model whose vision encoder produces the main state-of-the-art result.","marker":"[6]"},{"why":"Supplies UNI, the histopathology-pretrained encoder compared as a frozen feature extractor.","marker":"[12]"},{"why":"Supplies QUILT, the histopathology vision-language model included in the model comparison.","marker":"[10]"},{"why":"Supplies BiomedCLIP, the biomedical vision-language model included in the model comparison.","marker":"[8]"},{"why":"Provides the BCFC body-cavity-fluid cytology dataset used as one of the four benchmarks.","marker":"[25]"},{"why":"Provides the MLCC cervical lesion dataset used as one of the four benchmarks.","marker":"[26]"},{"why":"Provides the SIPaKMeD cervical cell dataset used as one of the four benchmarks.","marker":"[27]"}],"fun_headline_variants":["LoRA tuning lets CLIP beat cytology-specific models","CLIP with LoRA needs 70% data to top cytology benchmark","Light LoRA adaptation makes CLIP state-of-the-art on cytology","LoRA gives CLIP cytology edge over specialized transformers","General CLIP + LoRA outclasses cytology transformer with 62x fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["LoRA tuning lets CLIP beat cytology-specific models","CLIP with LoRA needs 70% data to top cytology benchmark","Light LoRA adaptation makes CLIP state-of-the-art on cytology","LoRA gives CLIP cytology edge over specialized transformers","General CLIP + LoRA outclasses cytology transformer with 62x fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1589,"prompt_tokens":863,"completion_tokens":726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":632}},"tokens_in":479,"tokens_out":726,"duration_ms":6531,"temperature":1.0,"reasoning_tokens":632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:39:28.211970+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A visual–language foun- dation model for pathology image analysis using medical twit- ter,","cited_arxiv_id":null,"evidence_quote":"Defines LoRA, the low-rank adaptation mechanism that carries all fine-tuning experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides CLIP, the general-purpose vision-language model whose vision encoder produces the main state-of-the-art result."},{"cited_title":"Deeppap: deep convolutional net- works for cervical cell classification,","cited_arxiv_id":null,"evidence_quote":"Supplies UNI, the histopathology-pretrained encoder compared as a frozen feature extractor."},{"cited_title":"Dausort and T","cited_arxiv_id":null,"evidence_quote":"Supplies BiomedCLIP, the biomedical vision-language model included in the model comparison."},{"cited_title":"Deep learning for computa- tional cytology: A survey,","cited_arxiv_id":null,"evidence_quote":"Provides the SIPaKMeD cervical cell dataset used as one of the four benchmarks."}],"review_version":1}