{"id":"b3e8bc44-1d45-40e0-b5a5-5998f8904c58","arxiv_id":"2504.12807","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A DenseNet201-encoded U-Net whose three hyperparameters are chosen by spider monkey optimization is reported to reach 96.16% accuracy and 95.63% Dice on SIPaKMeD Pap smear segmentation, but the evaluation protocol and baseline settings are not disclosed.","lead":"This paper combines a DenseNet201 image encoder with a U-Net decoder and uses spider monkey optimization to pick learning rate, batch size, and epoch count for segmenting Pap smear images. It reports high accuracy and Dice scores on a public cervical cell dataset, but the comparison baseline and optimization details are undisclosed, so the real gain is unclear.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No validation set or train/test split is described, and SMO's hyperparameter selection appears to be evaluated on the same data used for selection; the causal claim that SMO drives the reported large gains is therefore unverifiable.","rationale":"I agree with the reader's identification of the weakest assumption. The load-bearing condition is that the comparison between Dense-UNet201 with and without SMO is controlled: equivalent data splits, equivalent training budgets, and SMO's hyperparameters selected on validation data rather than on test data. None of these conditions is established in the manuscript. The paper reports no dataset split, no baseline hyperparameters, no SMO population size or iteration count, no validation-based selection protocol, and no repeated runs or confidence intervals. This is not an objection to metaheuristic hyperparameter search in general; it is a correctness risk specific to the reported experiment. The 10.9-point IoU improvement from Scenario 2 to Scenario 3 is large enough that distinguishing a genuine algorithmic contribution from test-set selection or an undertrained baseline is essential, and the supplied protocol cannot do so. The internal inconsistency in the learning-rate bounds further weakens confidence in the reported implementation details. With no code or formal verification supplied as independent support, the rejection remains appropriate.","tokens_in":11816,"tokens_out":3623,"duration_ms":38915,"concrete_test":"Require the authors to define a fixed train/validation/test split and rerun Scenario 3 with SMO optimizing only on training/validation, then freeze the selected hyperparameters and evaluate once on the test set. Repeat the Scenario 2 baseline with the same split, the same training budget, and disclosed baseline hyperparameters. If the test-set IoU gain from SMO falls materially below the reported 10.9-point jump, or if a random or grid search of the same budget matches it, the claim that SMO 'significantly enhanced' Dense-UNet201 is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SMO hyperparameter tuning raises Dense-UNet201 from 92.02% accuracy and 80.76% IoU in Scenario 2 to 96.16% and 91.63% in Scenario 3. This causal claim requires a controlled comparison: identical data splits, equal training budgets, and SMO's hyperparameters selected on a validation set rather than the test set. The paper provides none of these. Section 2.7 says metrics are reported 'on validation and test data,' but no train/validation/test split is defined anywhere, and Section 2.6 does not state how many SMO fitness evaluations were run or on which images. Section 2.2 reports only that Adamax was used for all models; the learning rate, batch size, and epoch count for the non-SMO baseline are not given, so the Scenario 2 model may simply be undertrained. All results are single-run point estimates without error bars, so the large jumps in IoU and Dice shown in Figure 7 could reflect test-set overfitting or random variation. There is also an internal inconsistency: Section 2.2 says the SMO learning-rate search range is 1e-5 to 1e-2 but later states '10^-5 to 10^-1,' a tenfold difference. The central claim therefore rests entirely on an uncontrolled, one-run comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dense-UNet201, a U-Net variant with a pretrained DenseNet201 encoder, for semantic segmentation of Pap smear images from the SIPaKMeD dataset. The authors combine this architecture with spider monkey optimization (SMO) for hyperparameter tuning and with an SMO-optimized PMD-CLAHE preprocessing step. Three scenarios are compared: no preprocessing, preprocessing only, and preprocessing plus SMO optimization of Dense-UNet201. The central claim is that SMO raises Dense-UNet201 from 92.02% accuracy, 80.76% IoU, and 88.18% Dice in the second scenario to 96.16% accuracy, 91.63% IoU, and 95.63% Dice in the third scenario, and that this improvement is attributable to SMO.","tokens_in":12132,"tokens_out":5397,"duration_ms":51604,"significance":"If the central claim were established, the paper would provide a useful case study of metaheuristic hyperparameter optimization for medical image segmentation and a systematic comparison of pretrained encoders in U-Net. The manuscript has strengths: it uses a public benchmark dataset, reports standard segmentation metrics, compares four architectures under the same preprocessing, and gives a reasonably detailed description of the SMO modifications for categorical and discrete variables. However, the main causal claim that SMO is responsible for the large performance gains is not supported by the experimental design as reported. The lack of a defined data split, the absence of baseline hyperparameters, and the single-run evaluation mean that the reported improvement could plausibly be due to selection bias, undertrained baselines, or random variation.","major_comments":[{"comment":"The manuscript never defines a training/validation/test split for the SIPaKMeD images, even though §2.7 states that metrics are reported on both validation and test data. Without a specified split and a rule that SMO hyperparameter selection uses only the validation set, the third-scenario results may have been selected directly on the test set, which would inflate the reported accuracy (96.16%), IoU (91.63%), and Dice (95.63%). Please specify the split, the number of SMO fitness evaluations, and which data were used at each stage of the optimization.","section":"§2.6–2.7, §3.3"},{"comment":"The comparison that supports the central claim is uncontrolled. For the Scenario-2 baseline Dense-UNet201, the paper does not report the learning rate, batch size, or number of epochs; §2.2 only states that Adamax is used for all models. If the baseline was trained with different or poorly chosen hyperparameters, the SMO improvement reported in Figure 7 could reflect under-training rather than the optimization algorithm. The baseline and SMO-tuned runs must use the same data split, training budget, initialization, and evaluation protocol.","section":"§3.2 vs. §3.3"},{"comment":"All results are single-run point estimates without error bars, confidence intervals, or significance tests. Given the large reported jumps in IoU and Dice, repeated runs with different random seeds are needed to establish that the differences are not due to random variation. At minimum, the authors should report the number of runs and the variance across runs.","section":"Figure 7, §3.3"},{"comment":"The learning-rate search range is stated inconsistently within the same section: the text first says the range is 1e-5 to 1e-2 and later says 10^-5 to 10^-1. This factor-of-ten discrepancy changes the search space and prevents reproduction of the SMO optimization. Please correct the inconsistency and state the exact range used.","section":"§2.2"},{"comment":"The probability update for categorical parameters in Eq. (9) is not clearly defined: the numerator uses fitness(LL_k) while the denominator sums fitness over all monkeys, and the relationship between the index update in Eq. (8) and this probability is unexplained. This makes the modified SMO difficult to reproduce and should be clarified.","section":"§2.5, Eq. (9)"}],"minor_comments":[{"comment":"Figure numbering is inconsistent: §2.3 says 'Fig. 3 illustrates the adjusted DenseNet-201 architecture' but Figure 3 is the U-Net diagram, and §2.6 also refers to Figure 2 for the simulation scenario, duplicating an earlier Figure 2. All figures should be renumbered and cross-checked.","section":"Figures"},{"comment":"The U-Net description gives an input size of 572×572×1, whereas the experimental protocol in §2.1 resizes all images to 256×256. Please clarify which input size was actually used.","section":"§2.4"},{"comment":"The notation in Eqs. (11) and (12) defines P as the ground-truth object and G as the predicted object, which is the reverse of the conventional assignment; please align the notation with the text or with standard usage.","section":"§2.7, Eqs. (11)–(12)"},{"comment":"Reference [35] is cited as the source of the SMO-optimized PMD-CLAHE preprocessing but appears to be a preprint by the same authors; please provide a persistent identifier and enough detail in this paper to make the preprocessing reproducible without relying on that reference.","section":"References"},{"comment":"There are several typographical errors, including 'bith size' in §2.5, 'DenseNe201' in §1, and inconsistent notation for powers of ten (10^-5 vs. 1e-5); these should be corrected in a final pass.","section":"Throughout"},{"comment":"The dice-loss expression in Eq. (13) appears to contain a typo: the second summation uses C_c where G_c is expected. Please verify the formula against the cited source.","section":"§2.7, Eq. (13)"}],"recommendation":"reject","confidential_remarks":"The central experimental comparison is uncontrolled, and the SMO selection protocol is not defined; these issues would require new experiments rather than a textual revision. I also note that the preprocessing method in [35] is the authors' own prior work and is cited without independent validation, which increases the risk of circularity in the overall pipeline evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is an incremental benchmarking paper — DenseNet201 encoder in U-Net, tuned by spider monkey optimization, tested on SIPaKMeD. The combination is not new in any interesting way; DenseNet encoders and SMO hyperparameter tuning both exist, and the paper offers only a rounding/mapping tweak for discrete hyperparameters. What it does well: the architecture table is clear, the metric definitions are standard, and the three-scenario comparison (raw, enhanced, enhanced+SMO) is an honest attempt to isolate preprocessing and tuning effects.\n\nThe problem is the central claim: SMO takes Dense-UNet201 from 92.02% accuracy / 80.76% IoU to 96.16% / 91.63%. That gain is presented without any evidence that the comparison is controlled. No validation split is described; the paper says metrics are reported \"on validation and test data\" but never defines the split or says which numbers belong to which. The non-SMO baseline hyperparameters (learning rate, batch size, epochs) are not reported, so the baseline may simply be undertrained — a 10-point IoU jump from tuning is suspicious, especially with no error bars or repeats. The preprocessing itself (SMO-tuned PMD-CLAHE) is the authors' own prior work and not independently validated, and the SMO fitness evaluations are not reported. There is also a small internal inconsistency: the learning-rate range is given as both 1e-5 to 1e-2 and 10^-5 to 10^-1.\n\nSo the added value is minor and the evidence for the headline claim is not there. The paper would need a proper train/validation/test split, equal training budgets for all models, multiple runs, and at least one non-SMO hyperparameter search baseline (random search would do) before the SMO claim can be taken seriously. I don't think this deserves referee time as is; it's a desk-reject-and-resubmit-with-data situation, not a fatal flaw in the idea. If the authors add the missing protocol, the result might become a useful benchmark for the Pap smear segmentation niche.\n\nMy recommendation: reject, but the rejection is about missing evidence, not a wrong method. A serious editor would send it back for major revision or desk reject with encouragement to redo the experiments properly.","headline":"A plain encoder-swap plus metaheuristic tuning, but the headline accuracy gain is unsupported because the comparison lacks a controlled baseline and a defined validation protocol.","tokens_in":12637,"tokens_out":1988,"would_cite":false,"duration_ms":20476,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A DenseNet201-encoded U-Net tuned by spider monkey optimization reaches 96.16% accuracy and 91.63% IoU on Pap smear segmentation.","keywords":["Pap smear","cervical cancer","semantic segmentation","DenseNet201","U-Net","Spider Monkey Optimization","SIPaKMeD","CLAHE"],"falsifier":"Train Dense-UNet201 without SMO using exactly the learning rate, batch size, and epoch count that SMO selected, on the same SMO-PMD-CLAHE enhanced data and the same train/validation/test split; if it also reaches roughly 96% accuracy and 91% IoU, the SMO-specific claim is falsified. A supporting check is to confirm the SMO hyperparameters were chosen on a validation set and to report results from multiple random seeds.","tokens_in":11647,"feed_emoji":"🔬","tokens_out":5857,"duration_ms":54368,"temperature":0.7,"pith_summary":"This paper proposes combining a DenseNet201 encoder with a U-Net decoder and using the spider monkey optimization (SMO) algorithm to choose the learning rate, batch size, and number of epochs, after also using SMO to tune a Perona-Malik diffusion plus CLAHE preprocessing step. The authors claim that, on the SIPaKMeD cervical cytology dataset, the optimized model segments cell images with 96.16% accuracy, 91.63% IoU, and 95.63% Dice coefficient. They further claim this beats standard U-Net, Res-UNet50, and Efficient-UNetB0 under the same preprocessing, and that the SMO tuning step accounts for a large final jump in performance. The work positions pretrained encoders, SMO-tuned image enhancement, and metaheuristic hyperparameter search as complementary routes to better cervical cell segmentation.","feed_headline":"Spider-monkey tuning lifts Pap smear segmentation to 91.6% IoU","feed_subtitle":"An SMO-tuned DenseNet-UNet with enhanced preprocessing beats three standard U-Net variants on cervical cell images.","key_machinery":"The central mechanism is Dense-UNet201 plus a modified spider monkey optimizer. DenseNet201's dense blocks concatenate features from all preceding layers in the encoder, which the paper argues improves gradient flow and feature reuse, while the U-Net decoder restores spatial detail through transpose convolutions aligned with encoder skip connections and a bottleneck that stays at 16×16×1024. The SMO component encodes each candidate solution as a triple of learning rate, batch size, and epoch count; continuous updates act on the learning rate, discrete updates round the epoch count to stay in [10,100], and categorical updates move an index mapped to the batch-size set. The swarm objective is the segmentation loss, and the optimizer's search is what the paper credits for the final performance gain.","core_discovery":"On the SIPaKMeD cervical cytology dataset, the paper claims that a segmentation network built by replacing the U-Net encoder with an ImageNet-pretrained DenseNet201, trained on images enhanced by an SMO-tuned hybrid Perona-Malik diffusion and CLAHE filter, and with learning rate, batch size, and epoch count selected by a modified spider monkey optimization algorithm, reaches 96.16% accuracy, 91.63% IoU, and 95.63% Dice coefficient. The SMO modifications round discrete epoch values and map categorical batch sizes to indices so the swarm search can handle mixed-variable hyperparameters. The paper further claims that these scores exceed those of standard U-Net, Res-UNet50, and Efficient-UNetB0 under the same preprocessing, and that the SMO step raises Dense-UNet201 from 92.02% accuracy, 80.76% IoU, and 88.18% Dice to the final numbers while reducing loss from 13.28% to 5.03%.","pith_inferences":["A natural extension is to use the SMO-selected hyperparameters as a fixed configuration and retrain Dense-UNet201 without the swarm loop, which would isolate how much of the gain comes from the search itself versus from the discovered hyperparameter values.","The same mixed-variable encoding could be applied to other architectural choices, such as decoder depth, loss weights, or augmentation settings, turning SMO into a general network-configuration search rather than a three-parameter tuner.","Because the preprocessing step improves all compared models, a promising follow-up is to test whether SMO-tuned PMD-CLAHE also helps other cervical cytology datasets and other stain-normalization pipelines.","The reported single-trial improvements would be more convincing with multiple seeds and confidence intervals, since deep learning runs can vary by several IoU points across seeds."],"forward_implications":["If the reported numbers reproduce, Dense-UNet201 becomes a strong baseline for cervical cell segmentation on SIPaKMeD, with a pretrained encoder adding roughly 22 percentage points of IoU over standard U-Net in the raw-data scenario.","The mixed-variable SMO encoding gives a template for applying swarm optimizers to deep learning hyperparameters that include categorical choices like batch size, not just continuous learning rates.","The SMO-tuned PMD-CLAHE preprocessing step improves every compared architecture, so the paper positions image enhancement and segmentation as jointly optimizable components.","The SMO tuning step is claimed to be responsible for the largest single gain, raising IoU by 10.87 percentage points and Dice by 7.45 percentage points over the preprocessed but untuned model."],"supporting_citations":[{"why":"Defines the U-Net architecture whose decoder and skip connections form the segmentation backbone.","marker":"[11]"},{"why":"Introduces the spider monkey optimization algorithm that the paper modifies for mixed hyperparameters.","marker":"[23]"},{"why":"Supplies the SIPaKMeD dataset and the ground-truth masks used for training and evaluation.","marker":"[28]"},{"why":"Defines DenseNet201, the pretrained encoder whose dense connectivity is the basis of Dense-UNet201.","marker":"[33]"},{"why":"Reports the SMO-tuned hybrid PMD-CLAHE preprocessing that the paper applies before segmentation.","marker":"[35]"},{"why":"Defines the Dice coefficient used as an evaluation metric.","marker":"[36]"},{"why":"Defines the categorical Dice loss used as the training objective.","marker":"[37]"}],"fun_headline_variants":["Monkey-tuned UNet hits 95.6% Dice on Pap smears","Spider monkey boosts Pap smear segmentation to 91.6% IoU","Dense-UNet plus spider swarm: 96% accuracy on cervical cells","SMO-optimized DenseNet201 UNet beats rivals on Pap smears","Swarm-tuned hybrid UNet nails 91.6% IoU for cervical cytology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the non-SMO Dense-UNet201 baseline was trained with reasonable, comparable hyperparameters, so the large gain in the third scenario is caused by SMO rather than by starting from a poorly tuned baseline.","fun_headline_variants_meta":{"raw":{"variants":["Monkey-tuned UNet hits 95.6% Dice on Pap smears","Spider monkey boosts Pap smear segmentation to 91.6% IoU","Dense-UNet plus spider swarm: 96% accuracy on cervical cells","SMO-optimized DenseNet201 UNet beats rivals on Pap smears","Swarm-tuned hybrid UNet nails 91.6% IoU for cervical cytology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3344,"prompt_tokens":986,"completion_tokens":2358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2248}},"tokens_in":602,"tokens_out":2358,"duration_ms":17511,"temperature":1.0,"reasoning_tokens":2248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:21:23.547200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Dense-UNet201 without SMO using exactly the learning rate, batch size, and epoch count that SMO selected, on the same SMO-PMD-CLAHE enhanced data and the same train/validation/test split; if it also reaches roughly 96% accuracy and 91% IoU, the SMO-specific claim is falsified. A supporting check is to confirm the SMO hyperparameters were chosen on a validation set and to report results from multiple random seeds.","supporting_citations":[{"cited_title":"Allahqoli et al., ‘Diagnosis of cervical cancer and pre -cancerous lesions by artificial intelligence: A systematic review’, Diagnostics, vol","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net architecture whose decoder and skip connections form the segmentation backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the spider monkey optimization algorithm that the paper modifies for mixed hyperparameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SIPaKMeD dataset and the ground-truth masks used for training and evaluation."},{"cited_title":"Khozaimi, I","cited_arxiv_id":null,"evidence_quote":"Reports the SMO-tuned hybrid PMD-CLAHE preprocessing that the paper applies before segmentation."}],"review_version":1}