{"id":"98f6fe6c-8923-476a-8a9e-1261df487c21","arxiv_id":"2504.17255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A 3D transformer-based segmentation framework quantifies, for the first time, temperature-dependent changes in sweat gland volume, length, and surface-to-volume ratio from OCT skin volumes.","lead":"Researchers trained a 3D deep-learning network to automatically find sweat glands in 3D skin scans made with optical coherence tomography. They used it to show that sweat gland volume, surface area, and length rise with skin temperature while the surface-to-volume ratio falls.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temperature-response claim rests on unvalidated segmentation generalization to 10°C/43°C OCT volumes, with no demonstrated correspondence of the same glands across conditions.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the segmentation model is applied to temperature-condition images without validation, and no registration confirms that the same physical glands are compared. The central claim—that sweat gland 3D morphology changes with temperature—depends entirely on the model's segmentations being unbiased across conditions. If the model's errors correlate with temperature, every Table II p-value is an artifact of segmentation, not biology. This is not a manufactured concern: the paper provides detailed segmentation validation on the training-domain data (Bland-Altman, correlation, external testing in Sections IV-A and IV-B) but deliberately omits equivalent validation for the temperature experiment in Section IV-D. The absence of registration is equally concrete: without matching individual glands across sessions, the 60-gland samples per condition could be drawn from different anatomical locations or different glands, and the observed differences would reflect population sampling rather than temperature response. I agree with the reader that a conditional accept is appropriate: the segmentation method appears sound and useful, but the physiological claim requires the proposed validation before it can be accepted. No change to the reader's verdict is needed because the stress-test confirms the same concern rather than uncovering a new one.","tokens_in":11678,"tokens_out":2775,"duration_ms":31899,"concrete_test":"Take the Section IV-D temperature-condition volumes and have two expert annotators independently label sweat glands in a sample of volumes from each temperature. Run the proposed model on the same volumes and compute per-temperature Dice, IoU, and volume-bias relative to the manual labels; a significant drop at 10°C or 43°C relative to 33°C would indicate temperature-dependent segmentation bias. In parallel, register each 10°C and 43°C volume to its corresponding 33°C baseline using rigid or deformable registration, identify matched individual glands by spatial overlap, and recompute Table II metrics only on matched glands. If the significant differences disappear in the matched-gland analysis, the reported morphological response is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-D applies the trained model to OCT volumes acquired at 10°C, 33°C, and 43°C and reports significant changes in volume, surface area, length, and S/V ratio. The model was trained on twelve OCT volumes collected without controlled temperature variation (Section III-A), and no validation of segmentation accuracy on the temperature-condition images is reported. If temperature alters image appearance—contrast, speckle, lumen size, or tissue motion—the network may systematically under- or over-segment as a function of temperature. Because every morphological metric in Table II is derived from model predictions, the central claim is only as strong as the assumption that segmentation remains unbiased across temperature conditions. Additionally, no registration step is described, so the 60 samples per condition are not shown to be the same physical sweat glands; comparing different gland populations across sessions could produce the same reported differences. Both issues independently threaten the physiological interpretation: the observed differences could reflect segmentation artifacts or sampling mismatch rather than genuine morphological response.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D transformer-based deep-learning framework for segmenting human skin sweat glands in optical coherence tomography (OCT) volumes, using a Swin-UNETR backbone with added ECA channel attention, DLK and DFF blocks in the decoder, and a hybrid BCE-Dice loss. The authors report strong segmentation performance in 4-fold cross-validation (Dice 0.8925, IoU 0.8062) and compare favorably with several state-of-the-art 3D segmentation methods. The main novelty claimed is the application of the trained model to OCT images acquired at 10°C, 33°C, and 43°C, reporting statistically significant changes in sweat gland volume, surface area, length, and surface-to-volume ratio, interpreted as a 3D morphological response to temperature. The paper concludes that the method enables real-time, non-invasive quantification of sweat gland morphology and could serve as a clinical tool for thermoregulation studies and dermatological diagnosis.","tokens_in":11887,"tokens_out":3600,"duration_ms":34980,"significance":"If the temperature-response results are valid, this work would provide the first non-invasive, quantitative 3D characterization of sweat gland morphological adaptation to thermal stimuli, with potential value for studying thermoregulatory disorders and for clinical OCT-based skin assessment. The segmentation contribution is competently evaluated: the authors use a standard cross-validation protocol, compare against multiple baselines, provide Bland-Altman agreement analysis for volume and surface area, and release a public data link. However, the central temperature claim rests on two unvalidated assumptions: that the segmentation model remains unbiased across temperature conditions, and that the same physical glands are compared across sessions. These assumptions are not tested in the manuscript, and the statistical analysis further inflates significance by treating glands within the same subjects as independent. The segmentation results alone are solid, but the headline finding of temperature-dependent morphology is not yet supported at the same standard.","major_comments":[{"comment":"The segmentation model trained on twelve OCT volumes acquired without controlled temperature variation is applied directly to OCT volumes collected at 10°C, 33°C, and 43°C, but no validation of segmentation accuracy on these temperature-condition images is reported. Temperature changes can alter OCT image contrast, speckle statistics, tissue motion, and gland lumen appearance, potentially causing systematic under- or over-segmentation that varies with temperature. Since every morphological metric in Table II is derived from model predictions, the temperature comparison is only valid under the untested assumption that segmentation bias is temperature-invariant. Please provide manual annotation of a subset of temperature-condition volumes and report segmentation Dice/IoU per temperature, or otherwise demonstrate that the model's performance does not systematically differ across conditions.","section":"§IV-D (Temperature study)"},{"comment":"The text states that OCT images of 'the same skin region' were acquired and that quantitative analyses were performed on 'the same glands' across conditions, but no registration or gland-matching procedure is described. Without a demonstrable spatial correspondence of individual glands between the 10°C, 33°C, and 43°C sessions, the reported differences could simply reflect comparing different gland populations sampled at each temperature. Please describe the co-registration method (e.g., landmark-based alignment, B-scan matching, or 3D volume registration) and how individual glands were traced across sessions, or alternatively present an analysis that explicitly accounts for gland identity, such as paired comparisons on matched glands.","section":"§IV-D (Same-gland matching)"},{"comment":"The statistical analysis reports p-values computed from 60 samples per temperature condition, but the data are nested: the 180 samples come from only five subjects, so multiple glands within a subject are not independent observations. Treating them as independent is a form of pseudo-replication that inflates significance; for example, the reported p = 2.56e-7 for volume would not survive a proper mixed-effects analysis with subject as a random effect. Please re-analyze the data using a linear mixed model or summarize at the subject level (n = 5 per condition), and report the adjusted p-values. The current Table II header (n=5) is also inconsistent with the text's '60 samples per condition'; clarify the exact structure of the data and what the reported statistics are computed over.","section":"Table II and §IV-D (Statistical analysis)"}],"minor_comments":[{"comment":"The hybrid loss weights (α = 0.9, β = 0.1) are stated without justification or sensitivity analysis; a small ablation study over these weights would strengthen the claim that the chosen balance is appropriate for the class-imbalance problem.","section":"Eq. (1) and §II-D"},{"comment":"The table header reports n=5 for each temperature condition, while the text reports 60 samples per condition; please clarify whether n refers to subjects and how the 60 gland samples are distributed across subjects and sessions, and specify the unit of analysis for the p-values.","section":"Table II"},{"comment":"The figure caption mentions the Kruskal-Wallis test, but the text does not state which statistical test produced the p-values in Table II; specify the test used for pairwise comparisons and whether any multiple-comparison correction was applied.","section":"Fig. 9 and §IV-D"},{"comment":"The external testing section presents two datasets without quantitative evaluation (e.g., Dice or IoU) because no manual ground truth is reported; state this limitation explicitly or provide qualitative-quantitative metrics if any annotations exist.","section":"§IV-B"},{"comment":"The data link is a Tianchi notebook URL without a persistent identifier; consider depositing the dataset in a permanent repository (e.g., Zenodo) to ensure long-term accessibility and reproducibility.","section":"Data availability"},{"comment":"The phrase 'for the first time' is used twice; given the prior OCT-based sweat gland segmentation works cited (e.g., refs [14-16]), the claim should be hedged and the specific novelty (quantitative temperature-response measurement) stated more precisely to avoid overclaiming.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The segmentation contribution appears sound and well evaluated, but the paper's headline temperature-response claim is currently supported only under untested assumptions about segmentation generalization across temperature conditions and gland correspondence across sessions. These are load-bearing issues that can likely be fixed with additional validation and a more appropriate statistical analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper delivers a usable 3D sweat-gland segmentation method from OCT and the first quantitative look at how gland volume, surface area, length, and S/V ratio change with temperature. The 3D morphological observations are new and worth taking seriously. But the temperature-response claim has a load-bearing validation gap, and the statistics as reported overstate the evidence.\n\nWhat the paper does well: the segmentation performance is credible — Dice 0.8925, IoU 0.8062 on the validation folds, with comparisons against Swin UNETR, nnU-Net, UNETR, V-Net, and 3D-ResUnet. External testing on two held-out datasets and Bland-Altman agreement on volume and surface area are the right checks. The data link is a plus, and I see no circular reasoning: segmentation is scored against manual annotations, and morphology is measured directly from the predictions.\n\nNow the soft spots. Section IV-D applies the model to OCT volumes at 10°C, 33°C, and 43°C and reports significant differences, but the model was trained on twelve volumes collected without controlled temperature. No validation of segmentation accuracy on the temperature-condition volumes is reported. If temperature changes contrast, speckle, or lumen appearance, the network could systematically under- or over-segment in a temperature-dependent way, which would produce exactly the kind of monotonic volume and surface-area shifts seen in Table II. This is the main issue. Second, the text says the same skin region was imaged, but I cannot find a registration step that proves the same physical glands are compared across conditions; without that, comparing different gland populations could create the reported differences. Third, the statistics treat 60 samples per condition as independent when they come from five subjects. The p-values are therefore optimistic; a mixed-effects model with subject as a random effect is needed. The S/V ratio interpretation is plausible but explicitly speculative, and the paper says so.\n\nThe architecture is an incremental combination of Swin UNETR, ECA at skip connections, and DLK/DFF blocks from D-Net. That is fine — not every paper needs a new block — but the novelty is in the application and the measurement, not the architecture.\n\nThis paper deserves a serious referee. I would send it to peer review and ask for: (1) segmentation validation at the three temperatures, even a small annotated subset; (2) a described registration or same-gland matching protocol; (3) subject-level mixed-effects statistics. If those are added, the physiological claim becomes solid. If they are not, the conclusion should be softened to 'preliminary.'","headline":"The segmentation pipeline is credible and the temperature-response observation is genuinely new, but the paper's central physiological claim currently rests on unvalidated segmentation at 10°C and 43°C and on pseudo-replicated statistics.","tokens_in":12390,"tokens_out":2853,"would_cite":true,"duration_ms":25806,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Human eccrine sweat glands measurably change their 3D shape with skin temperature: smaller at 10°C and larger at 43°C in OCT scans.","keywords":["optical coherence tomography","sweat gland segmentation","3D deep learning","temperature response","skin thermoregulation","transformer segmentation","morphometric analysis","eccrine sweat glands"],"falsifier":"Re-run the temperature protocol with manual annotations or an independent high-resolution imaging modality on the same physical glands across sessions, with explicit 3D registration. If the volume and length differences between 10°C and 43°C disappear, or if they track segmentation-artifact proxies such as contrast or false-positive mask volume rather than true gland boundaries, then the morphological-response claim is refuted.","tokens_in":11515,"feed_emoji":"🌡️","tokens_out":6859,"duration_ms":64347,"temperature":0.7,"pith_summary":"The paper tries to establish that human eccrine sweat glands change their three-dimensional shape measurably when skin temperature changes, and that this change can be seen live and non-invasively with optical coherence tomography (OCT). The authors build a transformer-based 3D segmentation network that isolates coiled sweat glands from OCT skin volumes, then apply it to scans of the same skin regions at 10°C, 33°C, and 43°C. They report that gland volume, surface area, and length increase significantly with temperature (for example, volume rises from $2.5776\\pm 0.4617$ to $3.6605\\pm 0.5395$ ($\\times10^{-4}\\,\\mathrm{mm}^3$)), while the surface-to-volume ratio falls. A sympathetic reader would care because, if true, this makes sweat-gland 3D morphology a real-time, non-invasive readout of thermoregulatory state and a quantitative benchmark for studying dermatological conditions.","feed_headline":"Sweat glands visibly swell and shrink with temperature","feed_subtitle":"A 3D segmentation model shows volume and surface area shift significantly across 10°C, 33°C, and 43°C.","key_machinery":"The load-bearing object is the segmentation network itself: a 3D transformer-based encoder-decoder that processes OCT volumes with a sliding window, uses shifted-window self-attention for long-range context, and adds a channel-attention block at skip connections and dynamic large-kernel and feature-fusion blocks in the deep decoder layers. It produces the binary masks of coiled eccrine glands from which volume, surface area, length, and surface-to-volume ratio are computed. The temperature-response claim rides entirely on this mask quality being unbiased across 10°C, 33°C, and 43°C images.","core_discovery":"On the paper's own terms, the central discovery is that short-term temperature variation changes the 3D morphology of human eccrine sweat glands in a statistically significant, directionally consistent way. The segmentation model, evaluated by 4-fold cross-validation, reaches a Dice coefficient of $0.8925\\pm0.0355$, exceeding the compared 3D baselines, and produces masks whose volume and surface-area measurements agree with manual annotations in Bland-Altman analysis. Applied to 180 sweat-gland samples at three temperatures, the model reports that at 10°C glands are constricted (volume $2.5776\\pm0.4617\\times10^{-4}\\,\\mathrm{mm}^3$), at 33°C intermediate ($3.2046\\pm0.4034$), and at 43°C dilated ($3.6605\\pm0.5395$); surface area and length follow the same ordering, and the surface-to-volume ratio decreases. The authors interpret this as a morphological adaptation favoring sweating efficiency at high temperature and metabolic exchange at low temperature.","pith_inferences":["Separating the gland lumen from the glandular wall in the 3D masks would attribute the observed volume increase to lumen dilation rather than tissue expansion; the paper's masks do not distinguish the two.","Repeated temperature cycles (for example 33°C to 43°C and back) with explicit gland registration could reveal whether the morphological change is fully reversible or shows hysteresis.","Applying the same protocol to pathological glands, such as those in bromhidrosis, could turn the method into a diagnostic or treatment-monitoring tool.","Adding more temperature levels would test whether the volume response is monotonic or saturating, linking gland geometry to thermoregulatory dynamics."],"forward_implications":["Gland morphometry becomes a real-time, non-invasive readout of thermoregulatory state rather than a biopsy-only assessment.","The reported parameter ranges give a quantitative baseline for normal glands, letting pathological states such as bromhidrosis and hypohidrosis be compared numerically.","OCT-based fingerprint and sweat-pore systems will need to control for skin temperature, because gland size affects the 3D pore and ridge geometry.","The surface-to-volume trend provides a concrete quantitative target for models of sweat gland function under thermal stress."],"supporting_citations":[{"why":"Supplies the transformer encoder backbone with shifted-window self-attention that the proposed network adapts for small, slender targets.","marker":"[33]"},{"why":"Provides the efficient channel-attention block inserted at skip connections to recalibrate channel features without dimensionality reduction.","marker":"[34]"},{"why":"Supplies the dynamic large-kernel and dynamic feature-fusion blocks used in the deep decoder layers to enlarge the receptive field.","marker":"[35]"},{"why":"Provides the 3D unsupervised enhancement preprocessing that clarifies gland boundaries before manual annotation and training.","marker":"[39]"},{"why":"Represents the earlier 3D CNN approach for sweat-gland segmentation that this work directly extends and improves upon.","marker":"[15]"},{"why":"Provides a strong self-configuring 3D segmentation baseline whose difficulty with spiral structures motivates the proposed architecture.","marker":"[41]"}],"fun_headline_variants":["Sweat gland volume swells 42% from 10°C to 43°C","3D deep learning reveals sweat glands change shape with temperature","Non-invasive OCT tracks sweat gland 3D morphology in real time","Sweat glands constrict at 10°C, dilate at 43°C","Deep learning model quantifies sweat gland 3D response to heat"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the segmentation network, trained on OCT volumes collected without controlled temperature variation, segments glands just as accurately in scans taken at 10°C, 33°C, and 43°C, so the reported group differences are biological shape changes rather than temperature-dependent segmentation artifacts; the paper does not report a separate validation of segmentation accuracy on the temperature-conditioned images or a registration step proving the same physical glands are compared across sessions.","fun_headline_variants_meta":{"raw":{"variants":["Sweat gland volume swells 42% from 10°C to 43°C","3D deep learning reveals sweat glands change shape with temperature","Non-invasive OCT tracks sweat gland 3D morphology in real time","Sweat glands constrict at 10°C, dilate at 43°C","Deep learning model quantifies sweat gland 3D response to heat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3047,"prompt_tokens":968,"completion_tokens":2079,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1977}},"tokens_in":584,"tokens_out":2079,"duration_ms":14152,"temperature":1.0,"reasoning_tokens":1977,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:44:32.769898+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the temperature protocol with manual annotations or an independent high-resolution imaging modality on the same physical glands across sessions, with explicit 3D registration. If the volume and length differences between 10°C and 43°C disappear, or if they track segmentation-artifact proxies such as contrast or false-positive mask volume rather than true gland boundaries, then the morphological-response claim is refuted.","supporting_citations":[{"cited_title":"Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer encoder backbone with shifted-window self-attention that the proposed network adapts for small, slender targets."},{"cited_title":"ECA-Net: Efficient channel attention for deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the efficient channel-attention block inserted at skip connections to recalibrate channel features without dimensionality reduction."},{"cited_title":"Toward grou nd-truth optical coherence tomography via three-dimensional unsupervised deep learning processing and data,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D unsupervised enhancement preprocessing that clarifies gland boundaries before manual annotation and training."},{"cited_title":"Sweat gland extraction from optical coherence tomography using convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Represents the earlier 3D CNN approach for sweat-gland segmentation that this work directly extends and improves upon."},{"cited_title":"nnU-Net: a self -configuring method for deep learning -based biom edical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides a strong self-configuring 3D segmentation baseline whose difficulty with spiral structures motivates the proposed architecture."}],"review_version":1}