Investigating the Visual Cues of CNNs for Vascular Segmentation: A Case Study in Microscopy and Fundus Imaging

Cesar Henrique Comin; Weslley dos Santos Silva

REVIEW 4 major objections 4 minor 35 references

Investigating the Visual Cues of CNNs for Vascular Segmentation: A Case Study in Microscopy and Fundus Imaging

T0 review · 4 major / 4 minor · reviewed 2026-07-31 · deepseek-v4-flash

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

Pith's one-line read This paper argues that CNNs segment blood vessels primarily from local pixel intensity, with an effective receptive field saturating near 20 pixels, and that shape cues alone are insufficient.

desk verdict Solid ERF and shape-cue experiments, but the intensity-vs-texture ranking is not established because the perturbation procedure preserves the intensity histogram. read the letter →

arxiv 2607.23371 v1 pith:UB6RLNZY submitted 2026-07-25 eess.IV cs.CVcs.LG

Weslley dos Santos Silva , Cesar Henrique Comin This is my paper · ORCID

classification eess.IVcs.CVcs.LG

keywords bloodvesselsegmentationvisualcuestexturebiasintensityeffectivereceptivefieldU-Netfluorescencemicroscopyfundusimaging

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

This paper tries to establish which visual cues convolutional networks actually use when segmenting blood vessels in fluorescence microscopy and retinal fundus images, and to quantify how much spatial context they need. Using per-patch pixel shuffling and intensity normalization, it argues that pixel intensity is more important than local texture, yet networks still score above 75% accuracy when both cues are removed, implying they exploit other statistical regularities. Training networks on contours and centerlines alone, it finds shape cues are insufficient (best Dice 54–57% versus 90% with full images), so networks do not reconstruct vessel geometry from boundaries. The paper also claims the effective receptive field saturates around 20 pixels across both datasets, a 32×32 patch reaches peak performance in microscopy, and only fundus images show modest gains from larger context. If correct, this means lightweight local-feature models are enough for many vessel-segmentation settings and that failures are better explained by local appearance than missing global shape.

What carries the argument

The identifying machinery is the cue-isolation protocol: patch-level pixel shuffling to destroy texture, patch-level mean and standard-deviation normalization to destroy absolute intensity, homogenization (both together), contour-only and centerline-only inputs to strip internal vessel appearance, and a systematic sweep of kernel sizes, dilation ratios, and pooling across 160 U-Net configurations to map theoretical versus effective receptive fields, with the ERF measured by gradient magnitudes. This protocol lets the authors decompose performance into separate contributions from intensity, texture, shape, and spatial context, and to locate the context budget beyond which further architectura

What would settle it

Train a U-Net on DRIVE with an architectural cap that forces the effective receptive field below 10 pixels, for example using all 1×1 convolutions with no pooling, and show it matches the base model's Dice; that would break the claim that roughly 20 pixels of context are needed. Alternatively, on VessMAP, show that 32×32 input patches yield a large Dice drop relative to full images once data augmentation or class balance is changed, revealing that the plateau is an artifact of the particular training setup rather than a true context budget.

Watch

Extended reading notes

Core claim

The central discovery is that these CNNs are local-intensity machines: segmentation decisions rest on a compact spatial window of about 20 pixels, and local contrast and pixel statistics dominate over texture as defined by spatial arrangement. The evidence: removing texture by shuffling pixels inside 9×9 patches costs only 0.1% classification accuracy, while removing intensity by normalization costs more; removing both still leaves above 75% accuracy; and supplying only contours or centerlines of vessels, without interior appearance, drops Dice from about 90% to 54–57%, with strong oversegmentation. The receptive-field experiments show performance plateaus near an average kernel size of 3 an

Load-bearing premise

The ranking of intensity over texture rests on the assumption that shuffling a 9×9 patch removes texture while leaving everything else intact, and that per-patch normalization removes intensity without destroying features the network uses; the paper's own result that accuracy stays above 75% even after both perturbations shows this separation is not clean.

Editorial extensions

If this is right

For microscopy datasets like VessMAP, a model with an effective field near 20 pixels and input crops of 32×32 reaches the same plateau as full-image models, so deeper or globally attentive architectures add little for this modality.
Dilated convolutions that enlarge the theoretical receptive field produce no consistent Dice gains, indicating the networks already operate with sufficient context at the base configuration.
Because shape cues alone yield around 54–57% Dice versus 90% with full information, vessel segmentation models should be expected to fail when local appearance is ambiguous, not because they lack global geometry.
For fundus images, global context helps modestly, so dataset-specific global structure such as the optic disk and retinal boundaries is part of what models use; architecture choices should account for this.
Networks retain above-chance patch classification after both intensity and texture perturbations, implying they rely on additional uncharacterized statistical features; identifying those features becomes a clear next step.

Reading between the lines

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

If the patch-level cue isolation transfers to whole-image segmentation, a practical shortcut is to audit new imaging domains with the same cheap probes—shuffle, normalize, and measure effective receptive field—before committing to large architecture searches.
The residual accuracy after homogenization suggests the network may be using per-patch variance, local contrast statistics, or edge orientation that survive shuffling; a testable next experiment is to explicitly ablate these statistics, for example by matching variances across classes, and check whether accuracy collapses.
The 20-pixel effective receptive field may be a property of the thin vessels in these datasets rather than of the task itself; applying the same protocol to high-resolution images with larger vessels should shift the saturation point, which would sharpen the claim.
The paper's proposed cue-specific attribution maps could be built directly from the measured performance deltas: saliency weights for intensity, texture, and shape are quantified by the three ablations, turning a qualitative bias diagnosis into a per-model score.

Share X Bluesky LinkedIn Reddit HN

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

Short version: this is a transparent, well-scoped empirical study with one clean and useful result — the ERF saturation around 20 pixels / 32×32 patch size for these two vessel datasets — and one load-bearing claim that does not survive close reading: the "pixel intensity > texture" ranking.

The receptive-field experiment is the strongest part. Sweeping kernel sizes, dilations, and pooling across 160 configurations, measuring ERF with the Luo et al. method, and getting a consistent saturation plateau is solid work. The patch-size experiment corroborates it nicely on VessMAP (saturation at 32×32) and honestly reports that DRIVE keeps improving with context. That finding is independent of the cue-isolation problem and is worth building on.

The shape-cue ablation is informative but weaker. The contour condition (Dc) is a legitimate probe: best Dice 54.4% vs ~90% with full information, with the predicted oversegmentation pattern. The centerline condition (Ds) is, as the authors admit, ill-posed — no network can infer vessel caliber from a centerline — so it adds little beyond confirming the trivial.

The soft spot is the texture/intensity isolation. The stress-test note is right: per-patch z-normalization removes mean and variance but preserves the intensity histogram; pixel shuffling preserves the same multiset of values. So D−i,−t does not remove "intensity" in any complete sense. The >75% accuracy on that condition is likely the network exploiting standardized intensity distribution shape, not some mysterious "other features." And the D−i vs D−t comparison is apples-to-oranges: normalization and shuffling destroy different kinds of information, so the observed accuracy drop does not license the claim that intensity matters more than texture. The paper's own framing ("intensity normalization is relevant mostly for patches containing only vessel or background pixels") hints at this, but the abstract and conclusion state the ranking unhedged.

All that said, the authors are unusually upfront about limitations, the code and hyperparameters are public, and the experiments are repeated with confidence intervals. The paper deserves a serious referee — a good reviewer would send it back for a revised cue-isolation methodology (e.g., replacing patch intensities with sampled noise matched only on low-order statistics) or for substantially toned-down claims. The ERF and shape results are publishable as a case study; the cue-ranking headline needs fixing first.

Referee Report

4 major / 4 minor

Summary. This paper investigates which visual cues drive CNN-based vascular segmentation, using fluorescence microscopy (VessMAP) and fundus photography (DRIVE) as case studies. Three experiment blocks are reported: (i) 9x9-patch classification after per-patch intensity normalization, pixel shuffling, or both, in order to isolate intensity and texture; (ii) segmentation from contour-only (Dc) and centerline-only (Ds) inputs in order to isolate shape; and (iii) a sweep of kernel size, dilation, and downsampling across 160 U-Net variants, plus training on WxW input patches, to measure theoretical and effective receptive fields. The authors report that intensity is more important than texture, that more than 75% patch accuracy survives removal of both, that shape cues alone yield Dice scores of about 54-57% versus roughly 90% with full information, and that performance saturates with an effective receptive field near 20 pixels (32x32 context for VessMAP), while DRIVE continues to benefit from larger contexts.

Significance. If established, the findings would support practical design recommendations: lightweight U-Nets with modest receptive fields are sufficient for these modalities, and XAI/saliency methods should be made cue-aware rather than generic. The experimental design is generally transparent: repeated runs (5x patch classification, 3x RF sweep), confidence intervals, public code, and explicit discussion of limitations. The ERF is measured with an external method rather than inferred from architecture alone, and the patch-size manipulation offers a clean test of context need. However, the interpretation of the patch-perturbation experiment is not clean because the 'removal' operations preserve first-order intensity statistics; this undermines the intensity-vs-texture ranking. The conclusion about shape is partly supported by the non-ill-posed Dc experiment but is conflated with the ill-posed Ds experiment. The paper is therefore a valuable and reproducible case study, but its headline interpretive claims need revision.

major comments (4)

[§III.B, Eq. (1), Fig. 6] The labels 'intensity removed' for D−i and 'texture removed' for D−t are not accurate as cue isolation. Eq. (1) subtracts the patch mean and divides by the patch standard deviation; it preserves within-patch relative intensity differences and the full shape of the normalized intensity histogram. Random shuffling preserves the intensity histogram exactly and removes only spatial order. Thus D−i,−t still contains complete first-order intensity statistics, and the reported >75% accuracy on D−i,−t is what a histogram-based classifier could achieve. The comparison between D−i and D−t conflates the type of perturbation (removing mean/variance vs. removing spatial order) with the semantic cue being removed. The headline 'pixel intensity is more relevant than texture' (Abstract, §V) rests on this comparison and is not established. I recommend adding a baseline classifier trained on per-patch int
[§III.C, Table I] The centerline dataset Ds is ill-posed by construction: vessel caliber cannot be recovered from a centerline, so a Dice ceiling below 100% is guaranteed regardless of model capacity. The manuscript acknowledges this in §III.C, yet the Abstract and §IV.B treat Ds as part of the evidence that 'CNNs struggle to extrapolate full vessel geometry from shape cues alone.' The contour dataset Dc is not ill-posed in the same way and provides the actual test of shape extrapolation; its low Dice (54.4% for W-Net) supports the claim. I recommend presenting Dc as the primary shape-only experiment and Ds as an extreme auxiliary condition, and explicitly stating that the Ds result is an expected lower bound. This would remove the misleading impression that 57.0% Dice on Ds is a meaningful 'best' shape-only result.
[§III.B/§IV.A] The intensity/texture conclusions are drawn from 9x9-patch classification on VessMAP only, but the Abstract and Conclusion state them as properties of vessel-segmentation networks in the evaluated modalities. Patch classification is not the same task as full-image segmentation, and the paper does not show that the residual features exploited on 9x9 patches (e.g., intensity histograms) transfer to segmentation. Because DRIVE was excluded from the patch experiments because its vessels are too thin, the conclusion 'for these modalities, pixel intensity is more critical than local texture' overstates the support. Please either restrict the claim to VessMAP patch classification or add a segmentation-level probe (e.g., apply the same perturbations to full images and measure Dice changes).
[§IV.C, Fig. 7-9] The RF sweep varies kernel size, which changes the number of trainable parameters as well as the receptive field. The manuscript acknowledges this in §IV.C ('Increasing the kernel size also adds more trainable parameters and, in turn, increases capacity'), but the following discussion still attributes performance gains to larger receptive fields. The patch-size experiment (Fig. 9) is a cleaner test of context need and does support the saturation conclusion for VessMAP, so this issue is not fatal. However, the manuscript should either control for parameter count (e.g., by adjusting width/depth) or explicitly state that the kernel-size results measure a joint capacity-plus-RF effect, and should adjust the text accordingly.

minor comments (4)

[§III.D, Fig. 9] DRIVE images are originally 584x565, but H is set to 512 in the patch experiment. The manuscript does not state whether images were resized or cropped before patchification. This should be specified, as it can affect the global-context conclusion for DRIVE.
[§IV.B] The statement that the shape results for DRIVE were 'similar' to VessMAP is not supported by any table or figure. Please report at least the DRIVE Dice values (or include them in an appendix) to substantiate the generalization claim.
[§IV.A, Fig. 6] The text reports losses relative to the base dataset (e.g., 'accuracy loss of 7.1%') but does not give the absolute accuracies of D, D−i, D−t, and D−i,−t. Please include the absolute values in the text or figure caption for reproducibility.
[§II] The related work is adequate, but the comparison with Loos et al. could note explicitly that the present RF sweep keeps depth fixed while varying kernel/dilation/pooling, which clarifies the difference in the manipulation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper's claims are empirical measurements with external methods, held-out test splits, and no fitted-results-relabeled-as-predictions.

full rationale

This is an empirical study rather than a derivation, and I find no step in which a predicted quantity is equivalent by construction to an input or fitted parameter. The ERF values are measured using the external methodology of Luo et al. [10] ('The ERF was measured using the methodology of [10]'), and Dice scores are reported on held-out test splits of standard datasets, so the central performance claims are not self-validating. The patch-classification and cue-isolation experiments in Section III.B are the closest candidates for concern, but they are construct-validity issues, not circularity: the definitions of 'intensity removal' (Eq. 1, z-normalization) and 'texture removal' (pixel shuffling) do not make the empirical result that D−i hurts more than D−t true by definition. The observation that D−i,−t retains >75% accuracy ('D−i,−t still provides enough information for the neural network to classify patches based on other characteristics, achieving more than 75% accuracy') is a limitation about what the perturbations remove, not a circular reduction. The one self-reference, the authors' own VessMAP dataset [27], is used as experimental input data, not as evidence for a claim, and so is not load-bearing. The paper also explicitly frames its findings as dataset- and architecture-specific observations rather than universal consequences ('the findings regarding the ERF size and the relative importance of shape vs. texture should be viewed as observations within a controlled case study'). No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no external benchmark is replaced by a self-citation. Score 1 reflects only the presence of a minor self-citation in the dataset description, which is not load-bearing.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

This is an empirical paper, so the ledger consists of experimental design choices rather than fitted model parameters. No new physical or model entities are introduced. The central numbers (ERF ≈ 20 px, >75% residual accuracy, Dice drops) are measurements, not fits, but their interpretation depends on the perturbation and sampling choices listed.

free parameters (3)

9×9 patch size = 9 × 9 px
Chosen by hand (§III.B) to fit entirely inside VessMAP vessels; 539 patches per class. All texture/intensity conclusions are conditional on this scale and on the single-modality sample.
Border fraction thresholds f = 75%, 50%
Hand-chosen (§III.B) to create border-containing patches; the claim that border benefit saturates (7.3% gain at f=75%, 4.5% at f=50%) depends on these two values.
ERF saturation value = ≈20 px
Informal visual read-off of Fig. 7 saturation, used in §IV.C/§V as the basis for recommending lightweight architectures; no formal criterion is stated.

assumptions (4)

domain assumption Pixel shuffling removes texture while leaving other statistics intact; per-patch normalization removes intensity level/scale without destroying other discriminative cues.
§III.B builds the entire texture-vs-intensity ranking on these two perturbation operations; the D−i,−t result (>75% accuracy) demonstrates residual signal, so the isolation is incomplete by the paper's own numbers.
domain assumption The gradient-distribution method of Luo et al. [10] measures the effective receptive field that the network actually uses.
All ERF and ERF/TRF claims in §IV.C inherit this external methodology; if ERF estimates are not faithful, the ≈20-pixel saturation claim is unsupported.
domain assumption Two datasets (VessMAP, DRIVE) and U-Net/W-Net variants are representative enough to support claims about CNN behavior for vascular segmentation.
The paper explicitly restricts itself ('should be viewed as observations within a controlled case study', §V); the generalization step is assumed, not demonstrated.
domain assumption Low Dice on the centerline-only Ds task evidences a CNN inability to use shape cues, despite the task being ill-posed.
§III.C labels Ds 'ill-posed'; using its low Dice to conclude CNNs 'cannot extrapolate vessel geometry' (§IV.B) assumes the failure is a network property rather than an information-theoretic impossibility.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Investigating the Visual Cues of CNNs for Vascular Segmentation: A Case Study in Microscopy and Fundus Imaging." pith.science (2026). https://pith.science/paper/UB6RLNZY

@misc{pith2026260723371,
  author       = {Pith},
  title        = {Pith review of: Investigating the Visual Cues of CNNs for Vascular Segmentation: A Case Study in Microscopy and Fundus Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UB6RLNZY}},
  note         = {Machine review of arXiv:2607.23371}
}

read the original abstract

Vascular segmentation is a standard procedure for clinical diagnosis, yet the specific visual features determining model decisions remain poorly understood. This paper investigates the visual cues Convolutional Neural Networks (CNNs) use to segment blood vessels across two distinct imaging domains: fluorescence microscopy and retinal fundus photography. We employ a series of experiments to quantify the influence of shape, texture, and receptive field on segmentation performance. First, we isolate texture and intensity by evaluating performance on patches subjected to pixel shuffling and normalization. Second, we assess global shape relevance by training models on sparse contours and centerlines. Lastly, we quantify the required spatial context by systematically varying the network's theoretical and effective receptive fields. Within the scope of the evaluated datasets, we found that pixel intensity is more relevant than texture, though networks maintain surprisingly high accuracy even when both cues are removed. Furthermore, CNNs struggle to extrapolate full vessel geometry from shape cues alone, typically relying on a relatively small effective receptive field of around 20 pixels, though global context provides a modest benefit for fundus images. While specific to the modalities studied, this methodology offers a quantitative foundation to audit and refine deep learning systems in vascular imaging.

Figures

Figures reproduced from arXiv: 2607.23371 by the authors.

**Figure 1.** shows examples of extracted patches. Visually distinguishing whether a patch belongs to a vessel or the background is difficult. This challenge motivates our quantitative analysis of a neural network’s ability to classify the patches. FIG. 1. Examples of 9 × 9 patches extracted from the VessMAP dataset. The first and second rows of images show, respectively, patches extracted from the blood vessels and the backgrou… view at source ↗

**Figure 2.** FIG. 2. Examples of 9 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3. Neural network used in the patch classification ex [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: FIG. 4. Example of images used for segmentation focused on [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: FIG. 5. U-Net model used for the receptive field experiments. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: FIG. 6. Average classification accuracy for different intensity [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: FIG. 7. Segmentation performance and receptive field sizes measured for U-Net architectures having different kernel sizes. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: FIG. 8. Segmentation performance and receptive field sizes measured for U-Net architectures having distinct dilation ratios. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: FIG. 9. Dice score as a function of patch size for the VessMAP [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 1 linked inside Pith

[1]

A review of machine learning methods for reti- nal blood vessel segmentation and artery/vein classifica- tion,

Muthu Rama Krishnan Mookiah, Stephen Hogg, Tom J MacGillivray, Vijayaraghavan Prathiba, Rajendra Pradeepa, Viswanathan Mohan, Ranjit Mohan Anjana, Alexander S Doney, Colin NA Palmer, and Emanuele Trucco, “A review of machine learning methods for reti- nal blood vessel segmentation and artery/vein classifica- tion,” Medical Image Analysis 68, 101905 (2021)

2021
[2]

All answers are in the images: A review of deep learning for cerebrovascular segmenta- tion,

Cheng Chen, Kangneng Zhou, Zhiliang Wang, Qian Zhang, and Ruoxiu Xiao, “All answers are in the images: A review of deep learning for cerebrovascular segmenta- tion,” Computerized Medical Imaging and Graphics 107, 102229 (2023)

2023
[3]

Hu- man treelike tubular structure segmentation: A compre- hensive review and future perspectives,

Hao Li, Zeyu Tang, Yang Nan, and Guang Yang, “Hu- man treelike tubular structure segmentation: A compre- hensive review and future perspectives,” Computers in Biology and Medicine 151, 106241 (2022)

2022
[4]

nnu-net: a self- configuring method for deep learning-based biomedi- cal image segmentation,

Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein, “nnu-net: a self- configuring method for deep learning-based biomedi- cal image segmentation,” Nature methods 18, 203–211 (2021)

2021
[5]

nnu-net revisited: A call for rigorous valida- tion in 3d medical image segmentation,

Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F Jaeger, “nnu-net revisited: A call for rigorous valida- tion in 3d medical image segmentation,” in International Conference on Medical Image Computing and Computer- Assisted Intervention (Springer, 2024) pp. 488–498

2024
[6]

Segment anything in medical images,

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang, “Segment anything in medical images,” Nature Communications 15, 654 (2024)

2024
[7]

Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead,

Cynthia Rudin, “Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead,” Nature machine intelligence 1, 206–215 (2019)

2019
[8]

Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural net- works,

Sajid Nazir, Diane M Dickson, and Muhammad Us- man Akram, “Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural net- works,” Computers in Biology and Medicine 156, 106668 (2023)

2023

Show all 35 references

[9]

This study was financed in part by the Coordena¸ c˜ ao de Aperfei¸ coamento de Pessoal de N ´ ıvel Superior - Brasil (CAPES) - Finance Code 001

for financial support. This study was financed in part by the Coordena¸ c˜ ao de Aperfei¸ coamento de Pessoal de N ´ ıvel Superior - Brasil (CAPES) - Finance Code 001
[10]

Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,

Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Bren- del, “Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,” in International conference on learning representations (2018)

2018
[11]

Understanding the effective receptive field in deep convolutional neural networks,

Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel, “Understanding the effective receptive field in deep convolutional neural networks,” Advances in neu- ral information processing systems 29 (2016)

2016
[12]

Survey of explainable ai techniques in healthcare,

Ahmad Chaddad, Jihao Peng, Jian Xu, and Ahmed Bouridane, “Survey of explainable ai techniques in healthcare,” Sensors 23, 634 (2023)

2023
[13]

Explainable deep learning models in medical image analysis,

Amitojdeep Singh, Sourya Sengupta, and Vasudevan Lakshminarayanan, “Explainable deep learning models in medical image analysis,” Journal of imaging 6, 52 11 (2020)

2020
[14]

Rethinking the image feature biases ex- hibited by deep convolutional neural network models in image recognition,

Dawei Dai, Yutang Li, Yuqi Wang, Huanan Bao, and Guoyin Wang, “Rethinking the image feature biases ex- hibited by deep convolutional neural network models in image recognition,” CAAI Transactions on Intelligence Technology 7, 721–731 (2022)

2022
[15]

Reducing texture bias of deep neural net- works via edge enhancing diffusion,

Edgar Heinert, Matthias Rottmann, Kira Maag, and Karsten Kahl, “Reducing texture bias of deep neural net- works via edge enhancing diffusion,” in ECAI 2024 (IOS Press, 2024) pp. 609–617

2024
[16]

Edges to shapes to concepts: Ad- versarial augmentation for robust vision,

Aditay Tripathi, Rishubh Singh, Anirban Chakraborty, and Pradeep Shenoy, “Edges to shapes to concepts: Ad- versarial augmentation for robust vision,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) pp. 24470–24479

2023
[17]

Shape prior is not all you need: Discovering balance between texture and shape bias in cnn,

Hyunhee Chung and Kyung Ho Park, “Shape prior is not all you need: Discovering balance between texture and shape bias in cnn,” in Proceedings of the Asian Con- ference on Computer Vision (2022) pp. 4160–4175

2022
[18]

Shape or texture: Understanding discriminative features in cnns,

Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bj¨ orn Ommer, Konstantinos G. Derpanis, and Neil Bruce, “Shape or texture: Understanding discriminative features in cnns,” in International Conference on Learn- ing Representations (2021)

2021
[19]

The origins and prevalence of texture bias in convolu- tional neural networks,

Katherine Hermann, Ting Chen, and Simon Kornblith, “The origins and prevalence of texture bias in convolu- tional neural networks,” in Advances in Neural Informa- tion Processing Systems, Vol. 33, edited by H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Cur- r...

2020
[20]

The estima- tion of the gradient of a density function, with appli- cations in pattern recognition,

Keinosuke Fukunaga and Larry Hostetler, “The estima- tion of the gradient of a density function, with appli- cations in pattern recognition,” IEEE Transactions on information theory 21, 32–40 (1975)

1975
[21]

Shape bias and robustness evalu- ation via cue decomposition for image classification and segmentation,

Edgar Heinert, Thomas Gottwald, Annika M¨ utze, and Matthias Rottmann, “Shape bias and robustness evalu- ation via cue decomposition for image classification and segmentation,” arXiv preprint arXiv:2503.12453 (2025)

2025 arXiv
[22]

On the influence of shape, texture and color for learning seman- tic segmentation,

Annika M¨ utze, Natalie Grabowsky, Edgar Heinert, Matthias Rottmann, and Hanno Gottschalk, “On the influence of shape, texture and color for learning seman- tic segmentation,” in ECAI 2025 , Frontiers in Artificial Intelligence and Applications, Vol. 413 (IOS Press, 2025) pp. ...

2025
[23]

Holistically-nested edge detection,

Saining Xie and Zhuowen Tu, “Holistically-nested edge detection,” in Proceedings of the IEEE international con- ference on computer vision (2015) pp. 1395–1403

2015
[24]

Convolutional neural networks rarely learn shape for semantic segmen- tation,

Yixin Zhang and Maciej A Mazurowski, “Convolutional neural networks rarely learn shape for semantic segmen- tation,” Pattern Recognition 146, 110018 (2024)

2024
[25]

Receptive field size as a key design parameter for ultrasound image segmen- tation with u-net,

Bahareh Behboodi, Maryse Fortin, Clyde J Belasso, Ru- pert Brooks, and Hassan Rivaz, “Receptive field size as a key design parameter for ultrasound image segmen- tation with u-net,” in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Bi- ology So...

2020
[26]

Understanding the influence of receptive field and network complexity in neural network-guided tem image analysis,

Katherine Sytwu, Catherine Groschner, and Mary C Scott, “Understanding the influence of receptive field and network complexity in neural network-guided tem image analysis,” Microscopy and Microanalysis 28, 1896–1904 (2022)

1904
[27]

Demystifying the effect of receptive field size in u-net models for medical image segmentation,

Vincent Loos, Rohit Pardasani, and Navchetan Awasthi, “Demystifying the effect of receptive field size in u-net models for medical image segmentation,” Journal of Med- ical Imaging 11, 054004–054004 (2024)

2024
[28]

A new dataset for measuring the performance of blood vessel segmentation methods under distribution shifts,

Matheus Viana da Silva, Nat´ alia de Carvalho Santos, Julie Ouellette, Baptiste Lacoste, and Cesar H Comin, “A new dataset for measuring the performance of blood vessel segmentation methods under distribution shifts,” PloS one 20, e0322048 (2025)

2025
[29]

Ridge-based vessel segmentation in color images of the retina,

J. Staal, M.D. Abramoff, M. Niemeijer, M.A. Viergever, and B. van Ginneken, “Ridge-based vessel segmentation in color images of the retina,” IEEE Transactions on Medical Imaging 23, 501–509 (2004)

2004
[30]

Understanding deep learn- ing requires rethinking generalization,

Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals, “Understanding deep learn- ing requires rethinking generalization,” in International Conference on Learning Representations (2017)

2017
[31]

Deep residual learning for image recognition,

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition (2016) pp. 770–778

2016
[32]

Building skeleton models via 3-d medial surface axis thinning algorithms,

Ta-Chih Lee, Rangasami L Kashyap, and Chong-Nam Chu, “Building skeleton models via 3-d medial surface axis thinning algorithms,” CVGIP: graphical models and image processing 56, 462–478 (1994)

1994
[33]

State-of-the-art retinal vessel segmentation with minimalistic models,

Adrian Galdran, Andr´ e Anjos, Jos´ e Dolz, Hadi Chakor, Herv´ e Lombaert, and Ismail Ben Ayed, “State-of-the-art retinal vessel segmentation with minimalistic models,” Scientific Reports 12, 6174 (2022)

2022
[34]

Very deep con- volutional networks for large-scale image recognition,

Karen Simonyan and Andrew Zisserman, “Very deep con- volutional networks for large-scale image recognition,” in International conference on learning representations (ICLR) (2015)

2015
[35]

ves- selfm: A foundation model for universal 3d blood vessel segmentation,

Bastian Wittmann, Yannick Wattenberg, Tamaz Ami- ranashvili, Suprosanna Shit, and Bjoern Menze, “ves- selfm: A foundation model for universal 3d blood vessel segmentation,” in Proceedings of the Computer Vision and Pattern Recognition Conference (2025) pp. 20874– 20884. APPEND...

2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [1]

A review of machine learning methods for reti- nal blood vessel segmentation and artery/vein classifica- tion,

Muthu Rama Krishnan Mookiah, Stephen Hogg, Tom J MacGillivray, Vijayaraghavan Prathiba, Rajendra Pradeepa, Viswanathan Mohan, Ranjit Mohan Anjana, Alexander S Doney, Colin NA Palmer, and Emanuele Trucco, “A review of machine learning methods for reti- nal blood vessel segmentation and artery/vein classifica- tion,” Medical Image Analysis 68, 101905 (2021)

2021

[2] [2]

All answers are in the images: A review of deep learning for cerebrovascular segmenta- tion,

Cheng Chen, Kangneng Zhou, Zhiliang Wang, Qian Zhang, and Ruoxiu Xiao, “All answers are in the images: A review of deep learning for cerebrovascular segmenta- tion,” Computerized Medical Imaging and Graphics 107, 102229 (2023)

2023

[3] [3]

Hu- man treelike tubular structure segmentation: A compre- hensive review and future perspectives,

Hao Li, Zeyu Tang, Yang Nan, and Guang Yang, “Hu- man treelike tubular structure segmentation: A compre- hensive review and future perspectives,” Computers in Biology and Medicine 151, 106241 (2022)

2022

[4] [4]

nnu-net: a self- configuring method for deep learning-based biomedi- cal image segmentation,

Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein, “nnu-net: a self- configuring method for deep learning-based biomedi- cal image segmentation,” Nature methods 18, 203–211 (2021)

2021

[5] [5]

nnu-net revisited: A call for rigorous valida- tion in 3d medical image segmentation,

Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F Jaeger, “nnu-net revisited: A call for rigorous valida- tion in 3d medical image segmentation,” in International Conference on Medical Image Computing and Computer- Assisted Intervention (Springer, 2024) pp. 488–498

2024

[6] [6]

Segment anything in medical images,

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang, “Segment anything in medical images,” Nature Communications 15, 654 (2024)

2024

[7] [7]

Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead,

Cynthia Rudin, “Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead,” Nature machine intelligence 1, 206–215 (2019)

2019

[8] [8]

Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural net- works,

Sajid Nazir, Diane M Dickson, and Muhammad Us- man Akram, “Survey of explainable artificial intelligence techniques for biomedical imaging with deep neural net- works,” Computers in Biology and Medicine 156, 106668 (2023)

2023