{"id":"3f6f76dc-9e3e-464a-85c7-901f573fe3cc","arxiv_id":"2505.15505","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-resolution CNN with 1.7M parameters achieves 91.28% accuracy on the SIPaKMeD cervical cell dataset, paired with a UNet-based multi-task segmenter, but the claimed progression-risk score is not validated.","lead":"This paper proposes a lightweight deep learning system for cervical cancer Pap smear analysis, with separate stages for cell segmentation, classification, and a probabilistic 'risk' score. The authors report accuracy comparable to much larger models, but the risk score is only a re-labeling of the same images, not a validated prognosis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The risk-assessment stage is a cross-sectional Gaussian class-probability estimate, not a progression predictor; the central prognosis claim is unsupported without longitudinal or external validation.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: the RA stage's posterior probability is treated as a progression likelihood without any longitudinal or external validation. I agree. The central claim in the abstract and conclusion is not merely a lightweight classifier; it is a three-stage workflow ending in prognosis. For that claim to hold, the quantity P(C_i|x) would have to be a valid predictor of malignant progression. The method as described cannot support this: it is fitted and evaluated on the same retrospective SIPaKMeD snapshots, and Table VI reports only class-separation accuracy of generic ML models on the extracted features. Neither the posterior nor the cosine-similarity threshold is checked against any outcome variable. This is an internal validity gap, not a disagreement with community consensus; the computed quantity is well-defined, but it is not the claimed quantity. I would additionally note that Table II contains a mathematical inconsistency (IoU 0.81, Dice 0.68 is impossible for binary segmentation because Dice = 2*IoU/(1+IoU)), which independently undermines trust in the segmentation evaluation. However, the decisive issue for the central claim is the unsupported prognosis framing. A clean longitudinal validation would settle the concern; without it, the paper's advertised risk-assessment contribution is not supported, and the verdict should remain rejection (or the paper should be revised to remove the prognosis claim and re-scope the contribution).","tokens_in":13464,"tokens_out":4125,"duration_ms":35959,"concrete_test":"On a longitudinal cervical-screening cohort with baseline Pap-smear cell images and follow-up histology (e.g., CIN2+ within 24 months), extract the 64-d MRF-DCN feature vector for each baseline cell, fit the Gaussian posteriors and the 0.65 cosine threshold on a training split, and evaluate discrimination of progressors versus non-progressors by AUC and calibration. If the AUC is near 0.5 or calibration is poor, the proposed RA stage does not function as a prognosis tool; if it performs well, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central advertised contribution is a probabilistic risk-assessment stage that 'predict[s] the likelihood of normal cells progressing to malignant states' (Abstract; Section I-B4; Section IV-E; Conclusion). What is actually computed is a posterior probability P(C_i|x) from multivariate Gaussian class-conditional densities fitted to 64-dimensional MRF-DCN features of the same SIPaKMeD images (Eqs. 1-5, Algorithm 1). That is a soft classifier over five cross-sectional cytology categories: it answers 'which class does this cell's feature vector resemble,' not 'will this normal cell progress to malignancy.' There is no longitudinal follow-up, no progression endpoint (e.g., CIN2+ on biopsy), no external cohort, and no temporal modeling anywhere in the manuscript. The threshold 'cosine similarity > 0.65 indicates a high probability of progression' is chosen from the same retrospective set (Section IV-E) and is never validated against outcomes. The high accuracy of the ML feature classifiers in Table VI only shows that the learned features separate the existing classes; it does not establish predictive validity for future malignant transformation. Because the title, abstract, contribution list, and conclusion all present RA/prognosis as a central output, this stage is not merely an optional extra: the central claim as stated fails unless the RA stage is reframed as in-sample class-similarity scoring or validated as a true prognostic marker. The Dice<IoU inconsistency in Table II is a separate reliability red flag, but the RA/prognosis gap is the load-bearing issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end pipeline for cervical cytology: cell segmentation by four standard models (UNet, LinkNet, FPN, PSPNet), classification by a new lightweight multi-resolution CNN (MRF-DCN, 1.7M parameters), and a \"risk assessment\" stage that attaches a probabilistic score to extracted features. The claims are: MRF-DCN achieves 91.28% accuracy on SIPaKMeD with far fewer parameters than VGG-19; a multi-task UNet reaches IoU 0.83 and 90% classification accuracy; and a Gaussian-based posterior over features predicts progression of normal cells to malignant states. The manuscript reports experiments on the SIPaKMeD dataset (and, in the conclusion, mentions Mendeley LBC) and concludes that the risk assessment can be used for prognosis.","tokens_in":13677,"tokens_out":2658,"duration_ms":24567,"significance":"If the claims were correct, the lightweight MRF-DCN would be a useful contribution to low-resource cervical screening, and the multi-task UNet would offer a compact unified segmentation-classification approach. The risk-assessment stage, however, is the advertised central novelty and would be clinically significant only if it actually measured progression. That claim is not demonstrated. The paper's strengths are the parameter-efficiency comparison against several standard CNNs and the explicit reporting of a multi-task loss formulation; however, the segmentation results contain a mathematical inconsistency, and the risk assessment is not validated as a longitudinal or prognostic tool.","major_comments":[{"comment":"For binary segmentation masks, Dice and IoU are related by Dice = 2·IoU/(1+IoU). With IoU=0.81, the UNet row should report Dice ≈ 0.90, not 0.68. The same inconsistency appears for all four models: LinkNet (IoU 0.55 vs DC 0.35; expected 0.71), FPN (0.75 vs 0.63; expected 0.86), and PSPNet (0.77 vs 0.68; expected 0.87). This indicates that at least one metric is computed incorrectly or from different masks, so the segmentation comparison as presented is not internally consistent and cannot support the claim that UNet and PSPNet are the most robust segmentation models.","section":"§IV-B, Table II"},{"comment":"The risk-assessment stage computes P(Ci|x) from multivariate Gaussian class-conditional densities fitted to 64-dimensional MRF-DCN features extracted from the same retrospective SIPaKMeD images. This is a soft classifier over five cross-sectional morphological categories; it answers 'which class does this feature vector resemble', not 'will this normal cell progress to malignancy'. The abstract, §I-B4, §IV-D, and §V all claim that the method 'predict[s] the likelihood of normal cells progressing to malignant states' and serves as a prognostic tool. No longitudinal follow-up, progression endpoint (e.g., CIN2+ or biopsy), external cohort, or temporal model is presented. The cosine-similarity threshold >0.65 in §IV-E is chosen post hoc from the same data and is not validated against any outcome. The high ML classification accuracies in Table VI only show that the extracted features separate the existing classes; they do not establish predictive validity for future transformation. This is a load-bearing mismatch between the advertised central contribution and what is actually computed.","section":"§IV-E, Eqs. (1)–(5) and Algorithm 1"},{"comment":"The ground-truth masks used for the segmentation experiments are never described. The text says 'Binary masks were generated for each patch' but does not state whether these masks come from manual annotation, automatic thresholding, or some other procedure. Without this information, the reported IoU/Dice values cannot be interpreted, and the bounding-box derivation from the predicted masks cannot be reproduced.","section":"§III-A"},{"comment":"The conclusion states that the risk-assessment method 'evaluated the SIPaKMeD dataset as well as the Mendeley LBC dataset', and §IV-E says the approach 'is effective on the Mendeley LBC Dataset ... as well'. However, no experimental results, tables, or quantitative comparisons for Mendeley LBC appear anywhere in the manuscript. This claim is unsubstantiated as written.","section":"§V and §IV-E"}],"minor_comments":[{"comment":"The multi-task loss in Eq. (14) depends on λ_seg and λ_cls, but Table III (the hyperparameter table for MRF-DCN) does not report these values, and the multi-task UNet section does not state its own hyperparameters or loss weights. The MTL result (IoU 0.83, accuracy 90%) is therefore not reproducible from the information given.","section":"§IV-C, Table III"},{"comment":"The column headers 'M W M W M W' are undefined and should be labeled (presumably macro/weighted averages). As written, the table is difficult to interpret.","section":"§IV-E, Table VI"},{"comment":"There are inconsistencies in class terminology: 'Dyskaryotic' and 'Dyskeratotic' are used interchangeably (e.g., Fig. 7 vs Table IV), and 'begets' in the abstract is nonstandard. These should be harmonized and edited.","section":"Throughout"},{"comment":"The class labels mix normal and abnormal categories; the statement that the model was trained to 'distinguish between the four classes' in §IV-D1 is inconsistent with the five-class SIPaKMeD setup described elsewhere. Please clarify the number of classes used in each experiment.","section":"§II, Table I"}],"recommendation":"reject","confidential_remarks":"The two most serious issues are the mathematically impossible Dice/IoU ratios in Table II and the unsupported leap from cross-sectional Gaussian posteriors to 'progression' and 'prognosis'. The first could be corrected by recomputing metrics, but the second is a fundamental mismatch between claim and method that would require new longitudinal or external outcome data, which is beyond a routine revision. The missing segmentation ground-truth description further weakens the work. In addition, the manuscript's self-described limitations (fixed resolutions, no external validation) do not address these core problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the classification half of this paper is a legitimate, if modest, empirical result; the risk-assessment half is a cross-sectional soft classifier being sold as a prognosis tool, and the segmentation metrics contain a mathematical impossibility. I would not cite it as is, but I would send it to review because the classifier deserves scrutiny, with the expectation of major surgery.\n\nWhat is actually new: MRF-DCN, a three-branch CNN with 1.7 million parameters, gets 91.28% accuracy on SIPaKMeD, a couple of points below the DeepCell result the paper itself cites. That is a real architecture, and the comparison to from-scratch baselines shows it is not obviously worse than much larger models. The multi-task UNet with a squeezed bottleneck is a minor tweak but a sensible one, and the paper's stated limitation about fixed input resolutions is honest.\n\nWhere it falls apart: the risk-assessment stage (Section IV-E, Eqs. 1-5) computes the posterior probability of a cell's feature vector under per-class Gaussians. That is a classifier, not a model of progression. There is no longitudinal follow-up, no CIN2+ endpoint, no external cohort. Calling it 'the likelihood of normal cells progressing to malignant states' in the abstract and conclusion is not supported by anything in the paper. The 0.65 cosine threshold is ad hoc and derived from the same retrospective set. The segmentation table is worse: Dice cannot be lower than IoU for the same binary prediction, yet Table II reports IoU 0.81 and Dice 0.68 for UNet, and similar contradictions for every model. That suggests the evaluation pipeline is unreliable, or at least carelessly reported. There is also no code or data, which compounds the problem.\n\nAlso, 'similar to state-of-the-art' is a stretch — DeepCell is 95.63% on the same dataset, and the paper's own related work notes MLNet outperforms existing models by roughly 2%. The headline accuracy is a modest contribution, not a breakthrough.\n\nWho this is for: someone building a lightweight cervical screening classifier might try the MRF-DCN architecture as a baseline. No one should use the risk-assessment stage as a prognostic marker until it is properly validated on longitudinal data.\n\nPeer-review recommendation: I would accept this for review, but only with a strong expectation of major revision: drop or re-validate the prognosis claims, fix or explain the segmentation metrics, and release code. If the Dice/IoU contradiction cannot be resolved, that alone justifies rejection.","headline":"A plausible lightweight classifier is buried under an unsupported prognosis claim and internally inconsistent segmentation metrics.","tokens_in":14319,"tokens_out":3282,"would_cite":false,"duration_ms":29429,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a 1.7M-parameter network classifies cervical cells at 91.28% accuracy, within 2-3% of much larger models, while a multi-task UNet reaches 0.83 IoU and 90% accuracy.","keywords":["cervical cancer","Pap smear","deep learning","image segmentation","cell classification","risk assessment","multi-task learning","lightweight CNN"],"falsifier":"Track a cohort of normal Pap smears forward in time and compare the model's risk scores with later biopsy-confirmed dysplasia or cancer; if high posterior probabilities do not predict progression, the risk-assessment claim fails. A simpler check is whether the risk score is just the classifier's confidence on a held-out set rather than a quantity tied to actual disease course.","tokens_in":13154,"feed_emoji":"🔬","tokens_out":7600,"duration_ms":62715,"temperature":0.7,"pith_summary":"This paper proposes a three-stage pipeline for Pap-smear screening: segment cell boundaries in whole-slide images, classify each cell into one of five categories, and score the risk that a normal cell is progressing toward malignancy. It reports that a compact three-branch network, MRF-DCN, classifies cells at 91.28% accuracy on the SIPaKMeD dataset with only 1.7 million parameters, roughly 85 times fewer than VGG-19 and within 2-3% of much larger models. It also reports a multi-task UNet variant that segments and classifies in one pass, reaching an IoU of 0.83 and 90% classification accuracy. The last stage fits a Gaussian probability model to the learned features to produce posterior probabilities for progression, which the authors offer as a prognostic tool.","feed_headline":"Cervical cell classification hits 91% with 1.7M parameters","feed_subtitle":"A lightweight multi-resolution CNN matches much larger networks on Pap smear images, making automated screening cheaper.","key_machinery":"The load-bearing object is MRF-DCN, a three-branch depthwise-separable convolutional network in which each branch consumes a different resolution of the same image and emits a 64-dimensional embedding; concatenation gives a 192-dimensional representation that is projected to 64 dimensions and classified by softmax. This multi-resolution design is what lets the model avoid resizing all inputs to one fixed scale and keeps the parameter count at 1.7 million. The second mechanism is the multi-task UNet, which appends a sequential classification layer to a squeezed bottleneck so segmentation and classification share one encoder and are optimized jointly. The risk score comes from a per-class multivariate Gaussian fit to the 64-dimensional features, with Bayes' theorem giving posterior probabilities and cosine similarity giving a second closeness measure.","core_discovery":"The authors' central claim is that all three screening tasks can be done with lightweight models without sacrificing accuracy. MRF-DCN processes each cell image at three resolutions (32x32, 64x64, 128x128) in parallel branches, fuses the three 64-dimensional embeddings into a 192-dimensional vector, and reduces it to 64 features before the final softmax; this configuration reaches 91.28% accuracy on the five SIPaKMeD classes. The multi-task UNet compresses its bottleneck and adds a classification head, training with a weighted sum of segmentation and classification losses, achieving IoU 0.83 and 90% accuracy simultaneously. The risk-assessment stage extracts the 64-dimensional features, models each class with a multivariate Gaussian, and uses Bayes' rule and cosine similarity to assign posterior probabilities that the authors interpret as the likelihood of normal cells progressing to koilocytotic or dyskaryotic states.","pith_inferences":["The risk probabilities are class-membership scores computed on a retrospective snapshot, not transition probabilities estimated from follow-up data; reading them as true progression likelihoods would require longitudinal validation.","The multi-resolution fusion principle should transfer to other cytology or histology datasets with variable cell sizes and magnifications, but the paper only tests it on SIPaKMeD.","A testable extension would be to replace the fixed 32/64/128 resolutions with a learned scale selection, addressing the fixed-scaling limitation the authors note."],"forward_implications":["A model with 1.7 million parameters and 91.28% accuracy could run on far less expensive hardware than the 144-million-parameter VGG-19 baseline, making automated Pap-smear triage more feasible in low-resource settings.","One multi-task network can produce both a segmentation mask and a five-class label in a single pass, removing the need to chain separate segmenters and classifiers.","If the posterior probabilities are trustworthy, cells flagged with high risk scores could be prioritized for pathologist review, compressing the workload of screening programs.","The framework's segmentation-to-bounding-box step turns masks directly into classification inputs, so the pipeline can go from whole-slide image to cell-level label automatically."],"supporting_citations":[{"why":"Supplies the SIPaKMeD dataset of 4049 cell images used for all classification, segmentation, and risk experiments.","marker":"[14]"},{"why":"Provides the UNet architecture that the segmentation and multi-task models are built on.","marker":"[7]"},{"why":"VGG-19 is the 144-million-parameter baseline against which the 85-fold parameter reduction is stated.","marker":"[39]"},{"why":"ResNet-101 and ResNet-18 are the compared baselines; ResNet-101 is the 25-fold larger reference.","marker":"[11]"},{"why":"MobileNetV3 is one of the lightweight baselines the MRF-DCN is compared with on the same split.","marker":"[10]"},{"why":"DeepCell reports 95.63% accuracy on SIPaKMeD, setting the state-of-the-art point the paper says MRF-DCN approaches within 2-3%.","marker":"[17]"},{"why":"Mendeley LBC dataset is used to claim the risk-assessment approach extends to L-SIL/H-SIL to squamous-cell carcinoma progression.","marker":"[15]"},{"why":"Segmentation models library is used to train UNet, LinkNet, FPN, and PSPNet for the binary segmentation comparison.","marker":"[36]"}],"fun_headline_variants":["Lightweight fusion net scores 91% on cervical cells","Multi-task UNet: 0.83 IoU, 90% accuracy on Pap","Bayesian model predicts cervical cancer progression risk","1.7M-param model rivals VGG-19 on Pap smears","Three tasks, one tiny CNN: cervical cancer screening"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The risk-assessment claim rests on the assumption that a cell's probability of belonging to a malignant class, computed from a statistical model fit to a retrospective snapshot, equals the real likelihood that the cell will progress to cancer; no follow-up data is used to test that equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight fusion net scores 91% on cervical cells","Multi-task UNet: 0.83 IoU, 90% accuracy on Pap","Bayesian model predicts cervical cancer progression risk","1.7M-param model rivals VGG-19 on Pap smears","Three tasks, one tiny CNN: cervical cancer screening"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3241,"prompt_tokens":947,"completion_tokens":2294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":563,"tokens_out":2294,"duration_ms":16166,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:16:58.144751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track a cohort of normal Pap smears forward in time and compare the model's risk scores with later biopsy-confirmed dysplasia or cancer; if high posterior probabilities do not predict progression, the risk-assessment claim fails. A simpler check is whether the risk score is just the classifier's confidence on a held-out set rather than a quantity tied to actual disease course.","supporting_citations":[{"cited_title":"Sipakmed: A new dataset for feature and image based classification of normal and pathological cervical cells in pap smear images,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIPaKMeD dataset of 4049 cell images used for all classification, segmentation, and risk experiments."},{"cited_title":"U-Net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the UNet architecture that the segmentation and multi-task models are built on."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"ResNet-101 and ResNet-18 are the compared baselines; ResNet-101 is the 25-fold larger reference."},{"cited_title":"Searching for mobilenetv3,","cited_arxiv_id":null,"evidence_quote":"MobileNetV3 is one of the lightweight baselines the MRF-DCN is compared with on the same split."},{"cited_title":"A deep neural network for cervical cell classification based on cytology images,","cited_arxiv_id":null,"evidence_quote":"DeepCell reports 95.63% accuracy on SIPaKMeD, setting the state-of-the-art point the paper says MRF-DCN approaches within 2-3%."},{"cited_title":"Liquid based- cytology pap smear dataset for automated multi-class diagnosis of pre- cancerous and cervical cancer lesions,","cited_arxiv_id":null,"evidence_quote":"Mendeley LBC dataset is used to claim the risk-assessment approach extends to L-SIL/H-SIL to squamous-cell carcinoma progression."},{"cited_title":"Segmentation models,","cited_arxiv_id":null,"evidence_quote":"Segmentation models library is used to train UNet, LinkNet, FPN, and PSPNet for the binary segmentation comparison."}],"review_version":1}