{"id":"63bf3c79-20db-4ef3-803e-086e721f5701","arxiv_id":"2501.12323","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A color-derived guiding map, concatenated to RGB patches, improves deep learning blood vessel segmentation in H&E oesophageal adenocarcinoma images by 1-5% Dice.","lead":"This paper adds a hand-made color map, the guiding map, as an extra input channel to deep learning models that find blood vessels in stained cancer tissue images. The extra channel helps the models score slightly higher on vessel segmentation, especially for tumor blood vessels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Improvement may be within noise: no significance tests, small 52-patch test set, overlapping error bars, and omitted baselines undermine the claimed gain.","rationale":"The reader correctly flags the missing significance testing and small test set, but their stated weakest assumption is the color-proxy reliability. I consider the statistical reliability more load-bearing: even if the guiding map perfectly highlights vessels, the paper's central claim is that it improves segmentation performance as measured by DSC/IoU. That claim cannot be evaluated without knowing whether the observed differences exceed chance. The tables show overlapping standard deviations; the 52-patch test set is small; and the omitted FCN/TransUnet guided-map results introduce selection bias. These issues affect internal validity, whereas the color-proxy concern is a generalizability limitation that the authors themselves acknowledge. Thus, the appropriate verdict remains CONDITIONAL, pending the requested significance analysis and full reporting.","tokens_in":106,"tokens_out":2611,"duration_ms":37251,"concrete_test":"For each architecture in Tables 1–3, compute the per-test-patch DSC difference between RGB and RGB+guiding-map predictions, then run a paired permutation test (or bootstrap 95% CI on the mean difference) using the 52 test patches. Also compute a Wilcoxon signed-rank test across the 5 folds. Report the number of folds (out of 5) in which the guiding-map input wins for each architecture. If the bootstrap CI includes zero for more than half the models, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that concatenating the guiding map improves segmentation—rests entirely on the DSC/IoU deltas in Tables 1–3. These deltas are small (0.3–3 percentage points overall, up to 5 for tumor BVs with EfficientUnet-B2) relative to the reported standard deviations (often 1–8 points). No significance test is reported. The test set contains only 52 patches from 3 WSIs, and the 5-fold protocol is described ambiguously, so it is unclear whether the error bars reflect fold-to-fold or patch-level variability. Moreover, the RGB+guiding-map rows for FCN and TransUnet were omitted from Tables 1–3 with the justification of 'poor performance'; this post-hoc exclusion means the claim is not tested on two of the seven architectures and may overstate consistency. Without a paired test at the patch or fold level, the observed improvements could be sampling noise, especially for the models with the smallest deltas.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a four-channel input for blood vessel segmentation in H&E-stained oesophageal adenocarcinoma whole-slide images, formed by concatenating a hand-crafted 'guiding map' to the RGB image. The map is generated by Gaussian blur, LAB conversion, Otsu thresholding of the A channel, luminosity scaling, morphological opening/closing, brightness scaling with the HSV V channel, and min-max normalization. Experiments compare RGB-only versus RGB-plus-guiding-map inputs across seven segmentation architectures (Basic U-Net, FCN, TransUnet, and four EfficientUnet variants) using DSC and IoU on a held-out set of 52 patches from 3 WSIs. The paper reports consistent small gains (about 1–4 percentage points) for most models, with the largest gain for tumour vessels using EfficientUnet-B2, but it reports no statistical significance testing, omits two models from the main comparison, and leaves the cross-validation protocol ambiguous.","tokens_in":5845,"tokens_out":6705,"duration_ms":64042,"significance":"If the gains are robust, the contribution is practically useful: a simple, architecture-agnostic input preprocessing step that can be dropped into existing U-Net-style pipelines, potentially reducing overfitting in small-data histopathology settings. The guiding map is derived from the input image by fixed image-processing operations rather than from labels or fitted parameters, so the RGB versus RGB-plus-guiding-map comparison is not circular by construction. Strengths of the paper include evaluation across several architectures and separate reporting for normal and tumour vessels. The main weakness is evidentiary: the test set is small (52 patches from 3 WSIs), error bars often overlap with the reported gains, and the missing statistical tests and omitted baselines make the central claim suggestive rather than conclusive.","major_comments":[{"comment":"The central claim that RGB-plus-guiding-map improves segmentation rests on DSC/IoU differences of roughly 0.2–4.4 percentage points. Many of these differences are smaller than or comparable to the reported standard deviations (e.g., EfficientUnet-B0 DSC 84.13±0.63 vs 85.16±1.64; EfficientUnet-B6 tumour DSC 85.81±1.34 vs 85.61±1.25). No paired significance test (e.g., Wilcoxon signed-rank or bootstrap on patches/folds) or confidence intervals are reported, yet Section 4 uses 'significantly increases' and Section 5 says 'significant improvements.' Because the test set contains only 52 patches from 3 WSIs, the observed deltas could be within sampling noise. Please add per-fold or per-patch paired tests, ideally with WSI-level clustering, for each model and input condition, and adjust the wording if the results do not reach significance.","section":"Section 4, Tables 1–3"},{"comment":"The data-splitting and cross-validation protocol is ambiguous and load-bearing. The text states 560 training/140 validation patches from 7 WSIs and 52 test patches from 3 WSIs, then says 'All experiments utilized 5-fold cross-validation.' It is unclear what is being cross-validated: whether the 5 folds are over the 700 training/validation patches, whether the same 52 test patches are reused across folds, and how the error bars in Tables 1–3 were computed (fold-to-fold or patch-level). The manuscript also does not report the number of patients or whether WSIs from the same patient appear in both training and test sets. Because neighbouring patches within a WSI are highly correlated, patch-level statistics can understate the true variability. Please specify the fold construction, the test-set usage across folds, the patient-level split, and the statistical unit for the reported means and standard deviations.","section":"Section 3.1 (Dataset and metrics)"},{"comment":"FCN and TransUnet are omitted from the RGB-plus-guiding-map condition 'because of their poor performance.' This post-hoc exclusion is not a valid experimental design for the paper's central claim: the claim that the guiding map improves segmentation is then tested on only five of seven architectures, and the reader cannot assess whether the guiding map harmed or helped the omitted models. If the results are excluded, the criteria for exclusion should be pre-specified, and the excluded numbers should still be reported in an appendix or supplementary table. Otherwise, the claim of consistent improvement is overstated.","section":"Section 3.3 and Table 1 caption"},{"comment":"The guiding-map pipeline has several hand-set components (Gaussian blur kernel k=3 and sigma=0, morphological kernel k=3, Otsu threshold) and no sensitivity analysis or control condition. In particular, there is no comparison with using a different single-channel input (e.g., the LAB L channel, HSV V channel, or a random noise channel) as the fourth channel, so it is not established that the specific guiding map, rather than simply an additional input channel, is responsible for the gains. Please add at least one control input and a brief sensitivity sweep of the main kernel sizes, or explicitly state that these parameters were fixed a priori and not tuned on the test set.","section":"Section 2.2 and Section 3.2"}],"minor_comments":[{"comment":"The title contains a word-splitting error: 'SEGMENTA TION' should be 'SEGMENTATION.'","section":"Title"},{"comment":"Reference [11] lists the TransUNet authors in corrupted order; the standard citation is Jieneng Chen, Yongyi Lu, Qihang Yu, et al.","section":"References"},{"comment":"The entry 'σ=0' for the Gaussian blur is ambiguous; in OpenCV, sigma=0 means the sigma is computed from the kernel size, so please state the effective sigma or report the intended value.","section":"Section 2.2, step 1"},{"comment":"The morphological kernel shape (square, disc, etc.) is not specified, which makes the guiding-map algorithm difficult to reproduce exactly.","section":"Section 2.2, step 4"},{"comment":"The row labeled 'Basic Unet [2]' appears to cite Glanzer et al. for the U-Net architecture; the appropriate reference for the basic U-Net is Ronneberger et al. [3].","section":"Table 1"},{"comment":"The text states the overall improvement is 'nearly 1-3% in DSC/IoU,' but Table 1 shows gains of about 4.2–4.4 percentage points for Basic U-Net; please rephrase to the actual observed range.","section":"Section 4"},{"comment":"The number of patients from which the 10 WSIs were obtained is not stated; please report it so readers can assess the generalizability of the results.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The submission is a short conference-style paper with a narrow claim. The main risk is statistical over-interpretation rather than methodological circularity, since the guiding map is a deterministic preprocessing of the input. Given the small test set, the missing significance tests, and the post-hoc omission of two architectures, I would not recommend rejection outright, but the authors need to address these statistical and protocol issues before the claim is acceptable. The work is potentially a useful practical contribution for computational pathology, but the evidence in the current form is not yet conclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper adds a hand-crafted color-based guiding map as a fourth channel to standard segmentation networks and reports consistent DSC/IoU gains on a private oesophageal H&E dataset. That is a legitimate, incremental contribution: the map is explicit, cheap, and architecture-agnostic, and the comparison is internally controlled because the same network and training protocol are used for both input variants. The qualitative examples look plausible, and the authors are honest that red cells are often absent or spilled, which limits the method's reach. I would not call this a rehash; the particular combination is new, and the idea of prepending a classical image-processing step to deep segmentation is worth knowing about.\n\nThe soft spots are real and all point in the same direction: the evidence is thin. The test set is 52 patches from 3 WSIs, the 5-fold cross-validation protocol is described ambiguously, and no significance tests are reported. Many of the reported gains are within one standard deviation of the baseline, so the central claim could be sampling noise, especially for the smallest deltas. The post-hoc omission of FCN and TransUnet from the RGB+GM condition because of \"poor performance\" is exactly the kind of exclusion that undermines the claim that the benefit is consistent across architectures. No code or data are released, so the reproducibility is limited.\n\nNone of this is fatal to the core idea. The guiding map is a sensible, low-cost input augmentation, and the fact that gains appear across several independent models and two BV subtypes is suggestive even if not conclusive. But the paper's language — \"significant improvements,\" \"greatly improves\" — goes beyond what the statistics support. With 52 test patches, a paired test at the patch or fold level is the minimum needed to believe the 1–3% numbers; reporting all baseline conditions, even failed ones, is also essential.\n\nWho benefits? Computational pathology researchers who work on vessel segmentation or who want a cheap trick to try in their own pipelines. It is a workshop-grade contribution, but it is not a throwaway: the comparison is controlled, the method is clearly described, and the authors flag the main limitation themselves. I would send this to peer review rather than desk reject, but I would reject the current version in review until the missing statistics and omitted baselines are addressed. If I were editing, I would make acceptance conditional on those additions plus code release for the guiding map generation.","headline":"A simple, honest, but statistically under-powered trick for vessel segmentation; worth a referee's time but the claimed gains are not yet established.","tokens_in":6376,"tokens_out":1614,"would_cite":false,"duration_ms":18764,"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":"Adding a hand-crafted guiding map to RGB input improves deep-learning blood-vessel segmentation in H&E-stained oesophageal adenocarcinoma whole-slide images.","keywords":["blood vessel segmentation","guiding map","H&E stained whole-slide images","oesophageal adenocarcinoma","deep learning","computational pathology","U-Net","color space thresholding"],"falsifier":"Compare RGB-only and RGB-plus-guiding-map performance on a test subset of vessels manually identified as having no red blood cells in their lumen (for example, drained or collapsed vessels); if the guided model does not match or exceed the RGB-only model on this subset, the color-prior claim is refuted for exactly the cases the paper says are hardest.","tokens_in":5473,"feed_emoji":"🩸","tokens_out":7357,"duration_ms":73781,"temperature":0.7,"pith_summary":"The paper claims that blood-vessel segmentation in H&E-stained oesophageal adenocarcinoma whole-slide images can be improved by giving a deep-learning segmentation network a hand-crafted color-derived guiding map as an extra input channel. The map is generated from the same RGB image through color-space conversion, thresholding, luminosity and brightness scaling, and morphology, then concatenated to the RGB channels to form a four-channel input. Across several U-Net style architectures, this extra channel raises Dice and IoU scores by roughly 1 to 3 percentage points overall, with larger gains of about 5 percentage points on tumor-associated vessels in one model. The authors argue this helps the models learn representative vessel features and reduces overfitting when labeled training data is scarce, which matters because manual vessel quantification in pathology is labor-intensive and rarely done in routine practice.","feed_headline":"Color-derived guide map lifts vessel segmentation on cancer slides","feed_subtitle":"On H&E oesophageal cancer slides, adding a red-content channel helps deep-learning models find tumor vessels.","key_machinery":"The central object is the guiding map, a hand-crafted single-channel prior generated from the input RGB image with no learned parameters. Its construction uses standard image-processing steps: Gaussian blur, conversion to LAB and HSV color spaces, Otsu thresholding of the LAB A channel to create a red-content heatmap, multiplication by the LAB L channel for luminosity scaling, morphological opening and closing to remove small objects and holes, multiplication by the HSV V channel for brightness scaling, and min-max normalization. Concatenated to the RGB channels as a fourth input, it supplies the segmentation network with a high-contrast hint about probable vessel locations. The paper's argument is that this extra channel guides the model to learn representative blood-vessel features, with the benefit growing in larger-capacity encoders where overfitting is more likely with limited training data.","core_discovery":"On H&E-stained oesophageal adenocarcinoma whole-slide images, concatenating a hand-crafted guiding map to the RGB input improves deep-learning blood-vessel segmentation compared with training the same networks on RGB alone. The guiding map is built by converting the RGB image to LAB and HSV color spaces, thresholding the LAB A (green-red) channel with Otsu's method, multiplying by the LAB L (lightness) channel, applying morphological opening and closing, multiplying by the HSV V (brightness) channel, and min-max normalizing to the range [0,1]. This single-channel map highlights regions with red blood cell content and is appended as a fourth input channel. The paper reports overall Dice/IoU improvements of about 1 to 3 percentage points across EfficientUnet variants, with the largest gain of nearly 5 percentage points on tumor blood vessels for the EfficientUnet-B2 model, and qualitatively smoother masks with fewer false positives. The authors attribute the benefit to the guiding map steering the network toward representative vessel features and mitigating overfitting in a label-limited setting.","pith_inferences":["Editorial extension: a control experiment that replaces the guiding-map channel with a random or edge-derived channel of matched statistics would test whether the gain comes from the specific red-content prior or merely from giving the network an extra input channel.","Editorial extension: the paper's own observation that red blood cells are inconsistently present implies the guiding map's advantage should shrink or disappear on vessels whose lumens contain no red blood cells; evaluating that subset separately would directly test the mechanism.","Editorial extension: the general recipe of extracting a color-opponent channel, thresholding it, scaling by luminosity and brightness, and concatenating it as an extra input might transfer to other H&E segmentation tasks whose target structures have a consistent color signature.","Editorial extension: because the reported standard deviations overlap across cross-validation folds, paired per-patch comparisons of RGB versus RGB-plus-guiding-map predictions would establish whether the improvement is consistent across test regions or driven by a few favorable patches."],"forward_implications":["The same segmentation architectures can be made more accurate for blood vessels by simply appending a fourth input channel, so the method is a low-cost addition to existing H&E-based vessel-segmentation pipelines.","The largest relative gains appear on tumor-associated vessels, the most irregular and clinically relevant class, suggesting the approach could support tumor-micro-environment studies that depend on vessel quantification.","Larger encoder models benefit more from the guiding map, which supports the paper's claim that the hand-crafted prior acts as a regularizer when training samples are limited or variable.","The reported smoother masks and reduced false positives would make downstream vessel-density measurements more stable if the effect replicates across datasets.","The method requires no architectural changes or additional staining, so it could be combined with future model improvements rather than competing with them."],"supporting_citations":[{"why":"Supplies the U-Net architecture on which the baseline and the EfficientUnet decoders are built.","marker":"[3]"},{"why":"Supplies the EfficientNet-B0/B2/B6 encoders whose RGB and RGB-plus-guiding-map inputs are compared.","marker":"[7]"},{"why":"Supplies the EfficientNetV2-M encoder used as one of the main tested models.","marker":"[8]"},{"why":"Provides the SCSE blocks added to the decoder layers of the EfficientUnet models.","marker":"[9]"},{"why":"Defines the U-Net-based vessel delineation baseline and appears as 'Basic Unet' in the results tables.","marker":"[2]"},{"why":"Provides the fully convolutional network baseline and motivates the use of H&E pathology images for microvessel analysis.","marker":"[5]"},{"why":"Supplies the TransUnet transformer-based baseline against which the guiding-map method is benchmarked.","marker":"[11]"}],"fun_headline_variants":["Color guide map boosts vessel segmentation on oesophageal cancer slides","Oesophageal cancer: color-derived map improves vessel segmentation accuracy","Adding a color-derived channel lifts vessel detection in H&E slides","Tumor vessel segmentation gets a boost from LAB-HSV guide map","Guide map from color spaces sharpens vessel segmentation on cancer slides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes a blood vessel can be seen and highlighted by its red content in the image, so vessels whose lumens contain no red blood cells or whose stain has shifted will get little or no guidance from the extra map.","fun_headline_variants_meta":{"raw":{"variants":["Color guide map boosts vessel segmentation on oesophageal cancer slides","Oesophageal cancer: color-derived map improves vessel segmentation accuracy","Adding a color-derived channel lifts vessel detection in H&E slides","Tumor vessel segmentation gets a boost from LAB-HSV guide map","Guide map from color spaces sharpens vessel segmentation on cancer slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000511,"raw_usage":{"total_tokens":2465,"prompt_tokens":907,"completion_tokens":1558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1473}},"tokens_in":523,"tokens_out":1558,"duration_ms":11545,"temperature":1.0,"reasoning_tokens":1473,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:16:17.305068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare RGB-only and RGB-plus-guiding-map performance on a test subset of vessels manually identified as having no red blood cells in their lumen (for example, drained or collapsed vessels); if the guided model does not match or exceed the RGB-only model on this subset, the color-prior claim is refuted for exactly the cases the paper says are hardest.","supporting_citations":[{"cited_title":"The HSV color space represents Hue, Saturation, and Value, while the LAB color space represents Lightness (L), and the op- ponent color channels A (green-red) and B (blue-yellow)","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net architecture on which the baseline and the EfficientUnet decoders are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EfficientNet-B0/B2/B6 encoders whose RGB and RGB-plus-guiding-map inputs are compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EfficientNetV2-M encoder used as one of the main tested models."},{"cited_title":"Dataset and metrics The dataset for this study was obtained from Imperial Col- lege London where H&E stained WSIs of oesophageal tis- sue were collected","cited_arxiv_id":null,"evidence_quote":"Provides the SCSE blocks added to the decoder layers of the EfficientUnet models."},{"cited_title":"Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images","cited_arxiv_id":"2501.12323","evidence_quote":"Defines the U-Net-based vessel delineation baseline and appears as 'Basic Unet' in the results tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fully convolutional network baseline and motivates the use of H&E pathology images for microvessel analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TransUnet transformer-based baseline against which the guiding-map method is benchmarked."}],"review_version":1}