{"id":"683f9835-3b6c-48db-b727-4c72a157d61c","arxiv_id":"2505.11735","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DenseNet-121 outperforms ResNet-50 in detecting brain cancer from bright-field microscopy images of thin brain tissue, reaching 88.35% test accuracy.","lead":"This study compares two deep learning models, ResNet-50 and DenseNet-121, for automatically detecting brain cancer in bright-field microscope images of thin brain tissue samples. DenseNet-121 classified test images more accurately than ResNet-50 (88.35% versus 82.12%), suggesting that densely connected networks may generalize better on small medical image datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TMA slide batch effects are a plausible confounder for the reported DenseNet-121 advantage; the paper provides no slide-level split or batch balance analysis.","rationale":"The paper is a straightforward empirical comparison with a patient-level split, which is commendable, and the McNemar test supports a real difference on this test set. My principal concern is that the comparison is only meaningful if the learned signal is biological. The manuscript's own physics rationale (intensity proportional to mass density) is imprecise, and the label trustworthiness rests on a brief mention of expert annotation. The most concrete missing evidence is slide-level independence. The reported confusion matrix and metrics are internally consistent, so I am not questioning arithmetic. The reader flagged batch artifacts as a key assumption; I agree that this is the load-bearing assumption, more so than the DenseNet-specific architecture claims. The requested slide-stratified split and batch-fingerprint test would settle it. If the gap persists under slide-level cross-validation, the paper's conclusion is substantially supported; if not, the central claim is an artifact of data handling. Because the paper is currently conditional on data availability and external validation, my critique does not change the verdict; it sharpens the condition.","tokens_in":5531,"tokens_out":5786,"duration_ms":61396,"concrete_test":"Request the slide identifier for every image and compute the class distribution per slide. Then perform a slide-stratified split: all cores from the same slide must remain in the same fold; retrain ResNet-50 and DenseNet-121 with the exact same two-stage protocol and multiple seeds (e.g., 5) per model. If the mean DenseNet minus ResNet test-accuracy gap is no longer significantly positive, or if accuracy drops by more than ~5 points relative to the reported 88.35%, the slide-level batch effect is a confounder. Additionally, train a small CNN to predict slide ID from held-out images; chance-level accuracy would suggest no batch fingerprint, while high accuracy confirms the model can exploit slide identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that dense connectivity itself causes DenseNet-121 to outperform ResNet-50 on limited medical data. That causal attribution requires that the images contain a cancer-specific tissue signature rather than slide-preparation artifacts. Section 2.1 describes a single commercial TMA: 24 cores per glass slide, 5-micron thickness, imaged on one Olympus BX61. All cores on a slide share the same staining, sectioning, and illumination conditions. The paper splits patients into train/validation/test (189/49/74) to prevent patient-level leakage, but it never reports how the TMA slides are distributed across these sets, whether cores from the classes are balanced within each slide, or whether any slide-level batch correction was applied. If cancer and non-cancer cores are concentrated on different slides, a classifier can achieve high accuracy by detecting slide-level intensity or stain variation rather than tumor morphology. This would inflate both the absolute accuracy and the DenseNet-ResNet gap (if DenseNet happens to be more sensitive to those cues), and the 'superior generalization of dense connectivity' conclusion in Sections 4-5 would not transfer to new clinical slides. The concern is concrete and testable; it is not a matter of consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript evaluates ResNet-50 and DenseNet-121 for binary classification of brain cancer from 2,931 bright-field transmission optical microscopy images of thin brain tissue. The data come from a commercial TMA, are split at the patient level into training/validation/test cohorts (189/49/74 patients), and are processed with a two-stage transfer-learning protocol, data augmentation, and early stopping. On the held-out test set, DenseNet-121 reaches 88.35% accuracy, 0.9614 precision, 0.8667 recall, and 0.9116 F1-score, while ResNet-50 reaches 82.12%, 0.9035, 0.8142, and 0.8563; the difference is reported as significant by McNemar's test (p = 0.004). The paper concludes that dense connectivity generalizes better on limited medical data.","tokens_in":1528,"tokens_out":1808,"duration_ms":57678,"significance":"If taken at face value, this is a modest but useful empirical benchmark on a relatively small single-site TMA dataset. The patient-level split is an appropriate safeguard against the most obvious data-leakage route, and the reported confusion-matrix numbers are internally consistent: 299 TP + 141 TN = 440 of 498 correct, matching 88.35% accuracy, and the precision, recall, and F1 values are consistent with those counts. The study does not propose a new method or theoretical result, and its value rests entirely on whether the learned signal is genuinely related to cancer morphology rather than slide-preparation or scanner artifacts, and on whether the DenseNet-versus-ResNet gap survives repeated splits and architectural ablations. These conditions are not currently established, so the significance of the headline claim is limited until the robustness analyses are supplied.","major_comments":[{"comment":"The manuscript never reports how the 24-core TMA slides are distributed among train, validation, and test sets, nor whether cancerous and non-cancerous cores are balanced within each slide. Because all cores on one slide share staining, sectioning, illumination, and scanner state, a classifier could separate classes by detecting slide-level batch differences rather than tumor morphology. This would inflate both the absolute accuracy and the DenseNet-ResNet gap if the two architectures have different sensitivities to those cues. Please add a slide-level breakdown, per-slide accuracy, and a leave-one-slide-out or slide-stratified cross-validation, or explicitly demonstrate that each slide contains both classes and that per-slide performance is stable. Without this, the 'superior generalization' conclusion in Sections 4 and 5 is not supported.","section":"Section 2.1 and Section 3.1"},{"comment":"The reported 95% confidence intervals are bootstrap intervals over the 498 test predictions, which quantify sampling uncertainty of the test set but not variance due to training initialization, augmentation order, or early stopping. With a single train/validation/test split, the 6.23% accuracy difference (p = 0.004 by McNemar) may not be robust. Please report results across multiple training runs with different random seeds and, ideally, across multiple patient-level splits, giving mean and standard deviation for accuracy, precision, recall, and F1-score. This is needed to establish that the DenseNet advantage is not an artifact of one favorable split or one favorable training run.","section":"Section 2.5 and Section 3.1"},{"comment":"The concluding attribution that 'These results underscore the value of dense connectivity for medical image analysis on limited datasets' is a causal claim about architectural design, but only two architectures are compared, and they differ in depth, width, connectivity pattern, and the details of which layers are fine-tuned. The observed gap could stem from any of these differences or from hyperparameter choices. To support the attribution, include an ablation that isolates dense connectivity (for example, a DenseNet variant with reduced connectivity or a controlled set of architectures matched in depth and width), or soften the conclusion to state that DenseNet-121 performed better on this specific dataset without making a general architectural claim.","section":"Section 5 and Section 4"},{"comment":"The study uses a single commercial TMA scanned on one microscope, and no external validation cohort is provided. The limitations paragraph acknowledges generalizability concerns, but the abstract and conclusions still state 'superior generalization' and point toward clinical translation. Please either add an external validation set from a different institution or TMA batch, or explicitly restrict the claims to the dataset at hand. As written, the wording overstates what a single-batch study can establish.","section":"Section 2.1 and Section 5"}],"minor_comments":[{"comment":"There is a typo: 'TAM samples' should be 'TMA samples'. Also, the citation markers [10,11] and [13,14,16] in this section do not align with the cited works in context; for example, reference [10] is the DenseNet paper and is not about TMA slide design. Please re-check all citation placements.","section":"Section 2.1"},{"comment":"The epoch counts are inconsistent: the text says early stopping halted DenseNet-121 at epoch 23 and ResNet-50 at epoch 17, while the caption of Figure 2 says 'over 36 epochs' and the Methods specify Stage 1 for 10 epochs and Stage 2 for 20 epochs. Please reconcile the numbers and make clear which training phase each figure corresponds to.","section":"Section 3.2 and Figures 2 and 3"},{"comment":"The caption describes panel (a) as loss and panel (b) as accuracy, but the panel labels appear reversed relative to the caption. Please verify the panel assignments.","section":"Figure 2"},{"comment":"The sentence stating that 'the transmission intensity of the microscope is proportional to the mass density of the samples' is repeated with slightly different wording in the same paragraph. Please consolidate and, if this proportionality is load-bearing for the biological interpretation, clarify how it translates into a learnable bright-field signature in the classification setting.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of physics.med-ph but the contribution is modest. The most important fix is the slide-level batch analysis, because without it the central DenseNet-versus-ResNet claim could be a batch-effect artifact. If the authors cannot provide slide-stratified validation, the conclusion about dense connectivity should be substantially weakened. I would not recommend rejection, because the reported metrics are internally consistent and the patient-level split is a genuine strength, but the paper needs the requested robustness experiments or a major change in the strength of its claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward comparison of two off-the-shelf CNNs on a new private bright-field brain TMA dataset. The reported metrics are internally consistent, and splitting at the patient level is good practice. But the paper's broader claim—that DenseNet-121's dense connectivity is the reason it wins on limited medical data—is not supported by the design, and the slide-level batch confounder is not addressed.\n\nWhat is actually new here is the specific benchmark: 498 test slices from a commercial brain TMA, with DenseNet-121 at 88.35% accuracy and ResNet-50 at 82.12%, p=0.004 by McNemar. The two-stage transfer learning protocol is described in enough detail to reproduce, the confusion matrix matches the stated metrics, and bootstrap CIs are given. That is a fair amount of care for a small empirical paper, and credit is due for the patient-level split.\n\nThe soft spots are real though. The biggest one is the one the stress-test flags: all cores on a TMA slide share the same staining, sectioning, and illumination, but the paper never reports how slides are distributed across train/validation/test or whether cancer and control cores are balanced within slides. If the classes are concentrated on different slides, the classifier can be reading slide-level brightness or stain variation rather than tumor morphology. That would inflate both the absolute accuracy and the DenseNet advantage, and it would gut the 'dense connectivity generalizes better' conclusion. This is not a nitpick; it is a testable design question that should have been answered in Section 2.1.\n\nOther issues are more minor: no multi-run variance, no external validation, no code or data release. The writing is sloppy in places—Section 2.1's intensity-to-mass-density sentence is garbled and repeated, and references [10], [11], and [16] are misapplied (one is a Pseudomonas paper, not brain tissue). The 'Limitations' paragraph is an incomplete sentence. None of these change the numbers, but they signal the manuscript needs another editing pass and a careful citation check.\n\nBottom line: this is a reasonable empirical data point for someone building a benchmark in small-sample bright-field histology, and it deserves a serious referee if the authors can add a slide-level batch analysis and temper the architectural claim. Without those, it is a single-dataset result, not a general lesson.","headline":"A credible but overinterpreted two-model benchmark on a private brain TMA dataset; the numbers are internally consistent, but the dense-connectivity conclusion outruns the single-dataset, single-split evidence.","tokens_in":6303,"tokens_out":3453,"would_cite":false,"duration_ms":33728,"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":"This paper argues that DenseNet-121, a convolutional network that reuses features through dense connections, outperforms ResNet-50 for detecting brain cancer in bright-field transmission microscopy images of thin tissue slices, reaching…","keywords":["brain cancer detection","bright-field microscopy","deep convolutional neural networks","DenseNet-121","ResNet-50","transfer learning","tissue microarray","medical image classification"],"falsifier":"Run the identical DenseNet-121 versus ResNet-50 protocol on images where the cancer labels are randomly permuted across slices from the same slide, or on an independent multi-center TMA set with the same two-stage training; if the 6.23% accuracy gap and the high precision disappear or shrink to noise, the claim that dense connectivity is systematically better on limited brain-tissue datasets is not supported.","tokens_in":5326,"feed_emoji":"🧠","tokens_out":7622,"duration_ms":68564,"temperature":0.7,"pith_summary":"This paper argues that DenseNet-121, a convolutional network whose layers reuse features through dense connections, is better suited than ResNet-50 for detecting brain cancer in bright-field transmission microscopy images of thin tissue slices. On a curated set of 2,931 images from tissue microarrays, DenseNet-121 reaches 88.35% test accuracy and a 0.9116 F1-score, against 82.12% and 0.8563 for ResNet-50, a gap the authors report as statistically significant (p=0.004, McNemar test). The authors read this as evidence that dense connectivity generalizes better on limited medical datasets, and they present their two-stage transfer-learning pipeline as a reusable blueprint for clinical bright-field workflows.","feed_headline":"DenseNet-121 beats ResNet-50 on brain cancer microscopy","feed_subtitle":"A head-to-head test on 2,931 brightfield tissue images finds dense connectivity yields 88.35% accuracy.","key_machinery":"The load-bearing object is DenseNet-121's dense connectivity: within each dense block, every layer receives the concatenated feature maps of all preceding layers, promoting feature reuse and gradient flow, and this is compared against ResNet-50's residual identity skip connections. The argument also depends on the two-stage transfer-learning protocol: first train a new classifier head on frozen ImageNet-pretrained features, then unfreeze the last two blocks and fine-tune with data augmentation, dropout, weight decay, learning-rate decay on plateau, and early stopping. This protocol is what the authors say lets both models converge cleanly on a 2,931-image dataset and lets the architectural difference show up as a statistically significant accuracy gap.","core_discovery":"The central claim is a head-to-head comparison: under identical two-stage transfer learning (frozen feature extractor trained for 10 epochs, then fine-tuning the last convolutional blocks for 20 epochs with augmentation, early stopping, and dropout), DenseNet-121 outperforms ResNet-50 on a held-out test set of 498 brain-tissue images. The reported numbers are 88.35% accuracy (95% CI ±1.28%), precision 0.9614, recall 0.8667, and F1 0.9116 for DenseNet-121, versus 82.12% accuracy (±1.75%), precision 0.9035, recall 0.8142, and F1 0.8563 for ResNet-50. The confusion matrix shows 299 true positives, 141 true negatives, 46 false negatives, and 12 false positives. The authors conclude that dense connectivity—connecting each layer to every other layer within dense blocks—improves feature propagation and gradient flow, and that this advantage is especially visible when training data are scarce.","pith_inferences":["The paper's architectural conclusion is drawn from a single dataset; on larger datasets or different tissue types the DenseNet advantage might shrink, since dense connections add memory and compute cost.","The claimed proportionality between transmission intensity and mass density could be tested directly by correlating model confidence or saliency maps with independent measurements such as quantitative phase or mass-density imaging.","A simple extension would be to switch from binary cancer detection to ordinal tumor grading; dense connectivity is naturally compatible with ordinal losses and the paper cites that direction.","One untested risk is slide-level batch effects: a leave-one-slide-out split, rather than the subject-level split used here, would reveal whether the model learns biology or slide-specific artifacts."],"forward_implications":["DenseNet-121 could serve as an automated pre-screening tool for brain cancer on bright-field biopsy slides, reducing manual review time.","The reported low false-positive rate (7.9%) implies fewer unnecessary follow-up procedures if such a model is used as a filter.","Architecture choice matters more on small medical datasets: dense feature reuse may be preferable to residual shortcuts when labeled images number only in the thousands.","The two-stage transfer-learning recipe (frozen head, then selective fine-tuning) can be carried over directly to other tissue types or stain protocols.","The statistically significant gap (p=0.004) supports prioritizing DenseNet-style connectivity in clinical AI pipelines for histopathology."],"supporting_citations":[{"why":"Supplies the ResNet-50 backbone that serves as the baseline in the head-to-head comparison.","marker":"[9]"},{"why":"Supplies the DenseNet-121 architecture whose dense connectivity is the method under test.","marker":"[10]"},{"why":"Earlier ResNet-versus-DenseNet comparison that frames the choice of the two backbones.","marker":"[12]"},{"why":"Documents the expert-characterized tissue microarray samples that form the dataset.","marker":"[13–15]"},{"why":"Describes the bright-field transmission scanning setup and the intensity-mass-density relation that motivates the learnable cancer signature.","marker":"[16]"}],"fun_headline_variants":["DenseNet-121 outperforms ResNet-50 on brain cancer images","88.35% accuracy: DenseNet-121 beats ResNet-50 in brain microscopy","Dense connectivity boosts brain cancer detection to 88%","DenseNet-121 wins over ResNet-50 for brain tissue screening"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the claim that the bright-field transmission intensity is proportional to tissue mass density, so that cancer-related structural changes produce a learnable optical signature in the images; if that link fails, or if the slice-level labels or the microscope slides carry batch artifacts, the reported accuracy and the DenseNet-versus-ResNet gap would not generalize beyond this dataset.","fun_headline_variants_meta":{"raw":{"variants":["DenseNet-121 outperforms ResNet-50 on brain cancer images","88.35% accuracy: DenseNet-121 beats ResNet-50 in brain microscopy","Dense connectivity boosts brain cancer detection to 88%","DenseNet-121 wins over ResNet-50 for brain tissue screening"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2183,"prompt_tokens":1018,"completion_tokens":1165,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1094}},"tokens_in":634,"tokens_out":1165,"duration_ms":9341,"temperature":1.0,"reasoning_tokens":1094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:48:49.617834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical DenseNet-121 versus ResNet-50 protocol on images where the cancer labels are randomly permuted across slices from the same slide, or on an independent multi-center TMA set with the same two-stage training; if the 6.23% accuracy gap and the high precision disappear or shrink to noise, the claim that dense connectivity is systematically better on limited brain-tissue datasets is not supported.","supporting_citations":[{"cited_title":"Huang, Z","cited_arxiv_id":null,"evidence_quote":"Supplies the ResNet-50 backbone that serves as the baseline in the head-to-head comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DenseNet-121 architecture whose dense connectivity is the method under test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier ResNet-versus-DenseNet comparison that frames the choice of the two backbones."},{"cited_title":"Sengupta, M","cited_arxiv_id":null,"evidence_quote":"Describes the bright-field transmission scanning setup and the intensity-mass-density relation that motivates the learnable cancer signature."}],"review_version":1}