Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images

Jiaqi Lv; Shan E Ahmed Raza; Stefan S Antonowicz

REVIEW 4 major objections 7 minor 25 references

Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Adding a hand-crafted guiding map to RGB input improves deep-learning blood-vessel segmentation in H&E-stained oesophageal adenocarcinoma whole-slide images.

desk verdict A simple, honest, but statistically under-powered trick for vessel segmentation; worth a referee's time but the claimed gains are not yet established. read the letter →

arxiv 2501.12323 v2 pith:YXN6MLTU submitted 2025-01-21 eess.IV cs.CV

Jiaqi Lv , Stefan S Antonowicz , Shan E Ahmed Raza This is my paper · ORCID

classification eess.IVcs.CV

keywords bloodvesselsegmentationguidingmapH&Estainedwhole-slideimagesoesophagealadenocarcinomadeeplearningcomputationalpathologyU-Netcolorspacethresholding

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

None 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.

Who 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.

Referee Report

4 major / 7 minor

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.

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 (4)

[Section 4, Tables 1–3] 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 3.1 (Dataset and metrics)] 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 3.3 and Table 1 caption] 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 2.2 and Section 3.2] 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.

minor comments (7)

[Title] The title contains a word-splitting error: 'SEGMENTA TION' should be 'SEGMENTATION.'
[References] Reference [11] lists the TransUNet authors in corrupted order; the standard citation is Jieneng Chen, Yongyi Lu, Qihang Yu, et al.
[Section 2.2, step 1] 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 2.2, step 4] The morphological kernel shape (square, disc, etc.) is not specified, which makes the guiding-map algorithm difficult to reproduce exactly.
[Table 1] 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 4] 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 3.1] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the guiding map is a fixed input-side transform, and the claimed improvement is an empirical comparison, not a reduction to the paper's own inputs.

full rationale

The paper's central claim is that concatenating a hand-crafted guiding map to RGB input improves blood vessel segmentation. The guiding map is generated from the input RGB image by a sequence of fixed image-processing operations (Gaussian blur, LAB/HSV conversion, Otsu thresholding of the A channel, luminosity and brightness scaling, morphology, and min-max normalization), as described in Section 2.2. It is not derived from the ground-truth labels, not fitted to the test set, and not defined in terms of the segmentation output. The comparison between RGB input and RGB+guiding-map input uses the same architectures, training protocol, and test set, so the measured DSC/IoU deltas are empirical outcomes rather than identities forced by construction. There are no load-bearing self-citations: the cited references (U-Net, EfficientNet, SCSE, TransUnet, etc.) are standard external works, and none is used to justify the guiding-map claim. The paper does not invoke a uniqueness theorem or an ansatz smuggled in via citation. Concerns about small sample size, missing significance tests, and omitted FCN/TransUnet rows in Tables 1–3 are statistical or reporting issues, not circularity. The derivation chain is self-contained with respect to the manuscript's own inputs and definitions, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a hand-crafted preprocessing pipeline with two hand-set kernel sizes, on the domain assumption that red blood cell color is a reliable vessel proxy, and on an unverified transfer-learning assumption. No new physical or ontological entities are introduced.

free parameters (2)

Gaussian blur kernel size and sigma = k=3, sigma=0
Hand-chosen in Section 2.2 step 1; no sensitivity analysis is reported for the smoothing parameters that shape the guiding map.
Morphological operation kernel size = k=3
Hand-chosen in Section 2.2 step 4; no sensitivity analysis is reported for the opening and closing kernel size.

assumptions (3)

domain assumption Blood vessels are detectable by red blood cell content in H&E slides, captured by the LAB A channel.
Section 2.1 states BVs are identified by red blood cells enclosed in an endothelial membrane, but also notes this content is inconsistent; the entire guiding map relies on this color proxy.
domain assumption ImageNet pretrained weights transfer useful features to H&E histology patches.
Section 3.2 initializes all models with ImageNet pretrained weights without comparing to from-scratch or histology-specific pretraining.
domain assumption The annotated ROIs represent WSI conditions and the data split avoids leakage.
Section 3.1 uses 28 ROIs from 10 WSIs and a 52-patch test set from 3 WSIs, but does not disclose patient-level grouping or stain normalization, leaving leakage risk unquantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images." pith.science (2026). https://pith.science/paper/YXN6MLTU

@misc{pith2026250112323,
  author       = {Pith},
  title        = {Pith review of: Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YXN6MLTU}},
  note         = {Machine review of arXiv:2501.12323}
}

read the original abstract

Blood vessels (BVs) play a critical role in the Tumor Micro-Environment (TME), potentially influencing cancer progression and treatment response. However, manually quantifying BVs in Hematoxylin and Eosin (H&E) stained images is challenging and labor-intensive due to their heterogeneous appearances. We propose a novel approach of constructing guiding maps to improve the performance of state-of-the-art segmentation models for BV segmentation, the guiding maps encourage the models to learn representative features of BVs. This is particularly beneficial for computational pathology, where labeled training data is often limited and large models are prone to overfitting. We have quantitative and qualitative results to demonstrate the efficacy of our approach in improving segmentation accuracy. In future, we plan to validate this method to segment BVs across various tissue types and investigate the role of cellular structures in relation to BVs in the TME.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

[1]

One of the hallmarks of cancer is inducing angio- genesis, a process through which tumors stimulate the for- mation of new blood vessels to sustain their growth

INTRODUCTION Blood vessels (BVs) in the Tumor Micro-Environment (TME) play a significant role in cancer progression and treatment re- sponse. One of the hallmarks of cancer is inducing angio- genesis, a process through which tumors stimulate the for- mation of new blood vessels to sustain their growth. These vessels provide essential nutrients and oxygen ...

work page
[2]

Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images

METHOD 2.1. The overall framework We propose a novel image processing algorithm for blood vessels (BVs) segmentation in H&E-stained images. BVs are typically identified by the presence of red blood cells enclosed within a thin endothelial membrane. However, the presence of red blood cells inside the membrane is inconsistent, as blood may be drained during...

work page Pith review arXiv 2025
[3]

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)

Color Space Conversions: Given an input RGB image xrgb of shape (H × W × 3), we apply Gaussian Blur to smooth the image (k=3, σ=0), then we convert it into two additional color spaces: HSV and LAB. 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) ...

work page
[4]

Thresholding with the A Channel: From xlab, we ex- tract the A channel, xa of shape (H × W × 1). Using Otsu’s thresholding method, we obtain the optimal thresh- old t that distinguishes red-colored pixels, then t is sub- tracted from xa to obtain xa′, any negative values are set to zero. This step generates a heat-map of red pixels

work page
[5]

Luminosity Scaling: xa′ is multiplied with the L channel from xlab to obtain xa′′, the aim is to scale the heat-map by luminosity

work page
[6]

The aim is to remove small objects and small holes

Morphological Operations: We apply morphological opening and closing operations on xa′′ (k=3). The aim is to remove small objects and small holes

work page
[7]

Brightness Scaling: We further enhance the contrast in brightness of xa′′ by multiplying it with the V channel from xhsv, which scales brightness linearly to obtainxa′′′

work page
[8]

Normalization: We apply Min-Max normalization to scale each pixel value in xa′′′ to the range of [0,1] ob- taining the guiding map which outlines potential blood vessels. Fig. 2 illustrates the steps involved in generating a guiding map. Fig. 3 shows examples of the resulting guiding maps

work page

Show all 25 references

[9]

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

EXPERIMENTS 3.1. 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. A total of 28 regions of interest (ROIs) were selected from 10 different WSIs with the aim to rep- resent ...
[10]

(we name these EfficientUnet). We incorporated Spatial and Channel Squeeze & Excitation (SCSE) blocks [9] into the decoder layers of the EfficientUnet models, these models were implemented using the PyTorch library [10]. Basic U- Net [3], FCN [5] and TransUnet (R50-ViT-B 16) [...
[11]

RESULTS AND DISCUSSION The test dataset contains both normal BVs and tumor BVs. Qualitative results, are shown in Fig.4 which demonstrate that the models trained with RGB + guiding map achieve superior segmentation quality compared to those trained with RGB input only. These m...
[12]

CONCLUSION AND FUTURE WORK We introduced a novel image processing algorithm to enhance deep learning-based segmentation of blood vessels (BVs) in H&E-stained oesophageal adenocarcinoma whole-slide im- ages. By generating high-contrast guiding maps and concate- nating them to R...
[13]

COMPLIANCE WITH ETHICAL STANDARDS The ethical approvals were obtained from Imperial College Tissuebank REC: 22/W A/0214, Project R23042-1A
[14]

JL is supported by the UK Engineering and Physical Sciences Research Council (EPSRC)

ACKNOWLEDGMENTS SA is supported by the NIHR Imperial Biomedical Research Centre, Cancer Research UK, and the Rosetrees Trust (ref: PGL23/100123). JL is supported by the UK Engineering and Physical Sciences Research Council (EPSRC). SEAR report financial support provided by the...
[15]

Hallmarks of cancer: the next generation,

Douglas Hanahan and Robert A. Weinberg, “Hallmarks of cancer: the next generation,” Cell, vol. 144, no. 5, pp. 646–674, March 2011

2011
[16]

Vessel de- lineation using u-net: A sparse labeled deep learning approach for semantic segmentation of histological im- ages,

Lukas Gl ¨anzer, Philipp Berens, Sonja N ¨orenberg, Sibylle Jablonka, and Britta Engelhardt, “Vessel de- lineation using u-net: A sparse labeled deep learning approach for semantic segmentation of histological im- ages,” Cancers, vol. 15, no. 15, pp. 3773, 2023

2023
[17]

U-net: Convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” CoRR, vol. abs/1505.04597, 2015

2015 arXiv
[18]

Fabnet : fea- ture attention-based network for simultaneous segmen- tation of microvessels and nerves in routine histology images of oral cancer,

Muhammad Moazam Fraz, S. A. Khurram, Simon Gra- ham, Muhammad Shaban, M. Hassan, A. Loya, and Nasir M. (Nasir Mahmood) Rajpoot, “Fabnet : fea- ture attention-based network for simultaneous segmen- tation of microvessels and nerves in routine histology images of oral cancer,” N...

2020
[19]

Microvessel pre- diction in h&e stained pathology images using fully con- volutional neural networks,

Fei Yi, Lei Yang, Shanshan Wang, Lei Guo, Chao Huang, Yao Xie, and Guozhu Xiao, “Microvessel pre- diction in h&e stained pathology images using fully con- volutional neural networks,” BMC Bioinformatics, vol. 19, no. 1, pp. 64, 2018

2018
[20]

Au- tomated quantification of blood microvessels in hema- toxylin and eosin whole slide images,

Azam Hamidinekoo, Anna Kelsey, Nicholas Trahearn, Joanna Selfe, Janet Shipley, and Yinyin Yuan, “Au- tomated quantification of blood microvessels in hema- toxylin and eosin whole slide images,” in Proceedings of the MICCAI Workshop on Computational Pathol- ogy, Manfredo Atzori...

2021
[21]

Efficientnet: Rethink- ing model scaling for convolutional neural networks,

Mingxing Tan and Quoc V . Le, “Efficientnet: Rethink- ing model scaling for convolutional neural networks,” CoRR, vol. abs/1905.11946, 2019

1905 arXiv
[22]

Efficientnetv2: Smaller models and faster training,

Mingxing Tan and Quoc V . Le, “Efficientnetv2: Smaller models and faster training,”CoRR, vol. abs/2104.00298, 2021

2021 arXiv
[23]

Recalibrating fully convolutional networks with spatial and channel ’squeeze & excitation’ blocks,

Abhijit Guha Roy, Nassir Navab, and Christian Wachinger, “Recalibrating fully convolutional networks with spatial and channel ’squeeze & excitation’ blocks,” CoRR, vol. abs/1808.08127, 2018

2018 arXiv
[24]

Segmentation models py- torch,

Pavel Iakubovskii, “Segmentation models py- torch,” https://github.com/qubvel/ segmentation_models.pytorch, 2019

2019
[25]

Transunet: Transformers make strong en- coders for medical image segmentation,

Chen Jieneng, Lu Yongyi Yu, Qihang, Luo Xiangde, Ehsan Adeli, Wang Yan, Lu Le, Alan L Yuille, and Zhou Yuyin, “Transunet: Transformers make strong en- coders for medical image segmentation,” arXiv preprint arXiv:2102.04306, 2021

2021 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

One of the hallmarks of cancer is inducing angio- genesis, a process through which tumors stimulate the for- mation of new blood vessels to sustain their growth

INTRODUCTION Blood vessels (BVs) in the Tumor Micro-Environment (TME) play a significant role in cancer progression and treatment re- sponse. One of the hallmarks of cancer is inducing angio- genesis, a process through which tumors stimulate the for- mation of new blood vessels to sustain their growth. These vessels provide essential nutrients and oxygen ...

work page

[2] [2]

Deep Learning Based Segmentation of Blood Vessels from H&E Stained Oesophageal Adenocarcinoma Whole-Slide Images

METHOD 2.1. The overall framework We propose a novel image processing algorithm for blood vessels (BVs) segmentation in H&E-stained images. BVs are typically identified by the presence of red blood cells enclosed within a thin endothelial membrane. However, the presence of red blood cells inside the membrane is inconsistent, as blood may be drained during...

work page Pith review arXiv 2025

[3] [3]

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)

Color Space Conversions: Given an input RGB image xrgb of shape (H × W × 3), we apply Gaussian Blur to smooth the image (k=3, σ=0), then we convert it into two additional color spaces: HSV and LAB. 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) ...

work page

[4] [4]

Thresholding with the A Channel: From xlab, we ex- tract the A channel, xa of shape (H × W × 1). Using Otsu’s thresholding method, we obtain the optimal thresh- old t that distinguishes red-colored pixels, then t is sub- tracted from xa to obtain xa′, any negative values are set to zero. This step generates a heat-map of red pixels

work page

[5] [5]

Luminosity Scaling: xa′ is multiplied with the L channel from xlab to obtain xa′′, the aim is to scale the heat-map by luminosity

work page

[6] [6]

The aim is to remove small objects and small holes

Morphological Operations: We apply morphological opening and closing operations on xa′′ (k=3). The aim is to remove small objects and small holes

work page

[7] [7]

Brightness Scaling: We further enhance the contrast in brightness of xa′′ by multiplying it with the V channel from xhsv, which scales brightness linearly to obtainxa′′′

work page

[8] [8]

Normalization: We apply Min-Max normalization to scale each pixel value in xa′′′ to the range of [0,1] ob- taining the guiding map which outlines potential blood vessels. Fig. 2 illustrates the steps involved in generating a guiding map. Fig. 3 shows examples of the resulting guiding maps

work page