arxiv: 2604.02977 · v1 · submitted 2026-04-03 · 💻 cs.CV

Recognition: no theorem link

Effect of Input Resolution on Retinal Vessel Segmentation Performance: An Empirical Study Across Five Datasets

Amarnath R

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:44 UTC · model grok-4.3

classification 💻 cs.CV

keywords retinal vessel segmentationinput resolutionthin vessel sensitivityDice scoreUNetfundus imagesmicrovascular segmentationwidth-stratified evaluation

0 comments

The pith

Reducing input resolution for retinal vessel segmentation drops thin vessel sensitivity by up to 15.8 points while Dice scores stay largely unchanged.

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

The paper tests how resizing fundus images to different resolutions affects UNet performance on retinal vessel segmentation across five datasets with native widths from 565 to 3504 pixels. All other training factors stay fixed while resolution changes, revealing that thin vessels (half-width under 3 pixels) lose critical detail when downsampled to subpixel sizes before the network even sees the data. A width-stratified sensitivity metric, derived from Euclidean distance transforms on native images, separates performance into thin, medium, and thick vessels and shows that overall Dice scores mask the problem because thick vessels dominate the count. For lower-resolution datasets the best thin-vessel results occur at or near native size, while high-resolution datasets peak after moderate downsampling to processed widths of 256–876 pixels. This pattern holds across all five sets and demonstrates that standard volumetric metrics are insufficient for judging microvascular accuracy.

Core claim

Training the same baseline UNet at multiple downsampling ratios on DRIVE, STARE, CHASE_DB1, HRF, and FIVES shows that thin vessel sensitivity improves monotonically toward the encoder’s effective range for high-resolution datasets, peaking between processed widths of 256 and 876 pixels, but is highest at native resolution for low-to-mid resolution datasets and degrades with any further downsampling; across all datasets aggressive downsampling reduces thin vessel sensitivity by up to 15.8 percentage points while Dice remains relatively stable.

What carries the argument

Width-stratified sensitivity metric that classifies vessels by native half-width (thin <3 pixels, medium 3–7 pixels, thick >7 pixels) using Euclidean distance transform estimates on the original images.

If this is right

Dice score alone cannot be trusted to evaluate segmentation of fine microvascular structures.
Optimal input resolution must be chosen according to each dataset’s native image width rather than fixed by GPU memory limits.
Irreversible information loss for thin vessels occurs before the network processes the image once vessels fall below pixel size.
High-resolution datasets benefit from moderate downsampling to match the network’s operating range, while low-resolution datasets do not.
Width-aware evaluation metrics are required to detect performance differences that standard overlap scores miss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Segmentation pipelines for any fine-structure task may need native or multi-scale inputs rather than uniform downsampling.
Clinical deployment of vessel segmentation tools could underperform on early microvascular changes if resolution choices are not dataset-specific.
Benchmark suites should adopt stratified sensitivity reporting to avoid over-optimism from thick-structure-dominated scores.
The same resolution-sensitivity trade-off is likely to appear in other domains that segment thin linear structures such as cracks or neural fibers.

Load-bearing premise

Keeping every training setting identical except input resolution fully isolates the effect of downsampling on thin-vessel detection without interference from changes in receptive field or feature scale.

What would settle it

Replicating the five-dataset experiment but upsampling all downsampled images back to native resolution before feeding them to the network; if thin vessel sensitivity no longer shows the reported drops, the claim that resolution-induced subpixel loss is the direct cause would be refuted.

read the original abstract

Most deep learning pipelines for retinal vessel segmentation resize fundus images to satisfy GPU memory constraints and enable uniform batch processing. However, the impact of this resizing on thin vessel detection remains underexplored. When high resolution images are downsampled, thin vessels are reduced to subpixel structures, causing irreversible information loss even before the data enters the network. Standard volumetric metrics such as the Dice score do not capture this loss because thick vessel pixels dominate the evaluation. We investigated this effect by training a baseline UNet at multiple downsampling ratios across five fundus datasets (DRIVE, STARE, CHASE_DB1, HRF, and FIVES) with native widths ranging from 565 to 3504 pixels, keeping all other settings fixed. We introduce a width-stratified sensitivity metric that evaluates thin (half-width <3 pixels), medium (3 to 7 pixels), and thick (>7 pixels) vessel detection separately, using native resolution width estimates derived from a Euclidean distance transform. Results show that for high-resolution datasets (HRF, FIVES), thin vessel sensitivity improves monotonically as images are downsampled toward the encoder's effective operating range, peaking at processed widths between 256 and 876 pixels. For low-to-mid resolution datasets (DRIVE, STARE, CHASE_DB1), thin vessel sensitivity is highest at or near native resolution and degrades with any downsampling. Across all five datasets, aggressive downsampling reduced thin vessel sensitivity by up to 15.8 percentage points (DRIVE) while Dice remained relatively stable, confirming that Dice alone is insufficient for evaluating microvascular segmentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Downsampling reduces thin vessel sensitivity by up to 15 points while Dice stays flat, but fixed UNet depth means receptive field scale changes too, so the mechanism is not fully isolated.

read the letter

The key point is that aggressive downsampling of fundus images reduces thin vessel sensitivity by as much as 15.8 percentage points on DRIVE while Dice scores stay stable. This holds across the five datasets and shows why standard metrics miss microvascular performance issues. The paper does a solid job comparing the same UNet setup on datasets with native resolutions from 565 to 3504 pixels. Introducing the width-stratified sensitivity, based on distance transform widths, lets them separate thin, medium, and thick vessels. The results are interesting: high-resolution datasets benefit from moderate downsampling, while lower-resolution ones prefer native size. That dataset dependence is the main new observation. The main limitation is that the fixed architecture means downsampling also enlarges the receptive field in original coordinates. Thin vessel pixels therefore see both fewer pixels and a broader context, so the sensitivity drop cannot be pinned solely on sub-pixel loss. No ablations with dilated layers or adjusted depths are shown to separate these effects. The abstract also lacks error bars or significance tests, which weakens the quantitative claims a bit. This work is aimed at people tuning retinal vessel segmentation models for clinical use, where missing thin vessels matters. A reader working on microvascular analysis would find the resolution guidelines practical. It is worth sending to peer review because the empirical pattern is clear and the metric addition is useful, even though the causal story needs tightening. Recommendation: Yes, put it through review and ask for receptive field controls in revision.

Referee Report

1 major / 3 minor

Summary. The manuscript reports an empirical study of how input resolution affects UNet performance on retinal vessel segmentation across five datasets (DRIVE, STARE, CHASE_DB1, HRF, FIVES). With all other hyperparameters fixed, the authors downsample images to multiple resolutions, introduce a width-stratified sensitivity metric (thin vessels: half-width <3 px; medium: 3–7 px; thick: >7 px) derived from native-resolution Euclidean distance transforms, and show that aggressive downsampling reduces thin-vessel sensitivity by up to 15.8 percentage points (DRIVE) while Dice scores remain relatively stable. High-resolution datasets improve at intermediate processed widths (256–876 px); lower-resolution datasets perform best near native resolution.

Significance. If the empirical patterns hold after addressing confounds, the work supplies concrete evidence that standard Dice scores can mask clinically important losses in microvascular detection and offers practical guidance on resolution selection for fundus segmentation pipelines. The multi-dataset scope and width-stratified evaluation constitute useful additions to the retinal-image-analysis literature.

major comments (1)

[§3 and §4] §3 (Experimental Setup) and §4 (Results): The fixed-depth UNet architecture implies that downsampling the input enlarges the receptive field relative to native fundus scale. Thin-vessel sensitivity drops therefore conflate sub-pixel information loss with changes in spatial context available to the encoder. The width-stratified metric, computed on native-resolution distance-transform widths after upsampling predictions, does not isolate these mechanisms; no ablation (dilated convolutions, variable encoder depth, or resolution-aware padding) is reported that would hold receptive-field size constant while varying only pixel density. This directly weakens the central attribution of performance change to irreversible pre-network information loss.

minor comments (3)

[Abstract and §4] Abstract and §4: Concrete quantitative differences are reported without error bars, standard deviations across random seeds, or statistical significance tests, leaving the reliability of the 15.8 pp DRIVE drop unclear.
[§3] §3: The exact set of tested input resolutions and the corresponding processed widths for each dataset should be tabulated for reproducibility.
[§4] §4: Clarify the interpolation method used when upsampling network outputs back to native resolution before metric computation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our empirical study. We address the major comment below and will incorporate clarifications into the revised manuscript.

read point-by-point responses

Referee: [§3 and §4] §3 (Experimental Setup) and §4 (Results): The fixed-depth UNet architecture implies that downsampling the input enlarges the receptive field relative to native fundus scale. Thin-vessel sensitivity drops therefore conflate sub-pixel information loss with changes in spatial context available to the encoder. The width-stratified metric, computed on native-resolution distance-transform widths after upsampling predictions, does not isolate these mechanisms; no ablation (dilated convolutions, variable encoder depth, or resolution-aware padding) is reported that would hold receptive-field size constant while varying only pixel density. This directly weakens the central attribution of performance change to irreversible pre-network information loss.

Authors: We agree that our fixed-depth UNet design means receptive-field size (in native fundus coordinates) changes with input resolution, so the thin-vessel sensitivity reductions reflect the joint effect of sub-pixel information loss and altered spatial context. This design choice was deliberate: it mirrors the standard practice in which practitioners resize images to fit GPU memory or batch constraints while leaving the network architecture unchanged. The width-stratified metric, evaluated after upsampling predictions to native resolution, still demonstrates that aggressive downsampling harms microvascular detection even under these typical conditions, and that Dice scores alone mask the loss. We will revise the discussion in §5 to explicitly acknowledge the confound and note that isolating the two mechanisms would require additional ablations (e.g., dilated convolutions or depth-varying encoders) that lie outside the scope of the present empirical survey. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical study with no derivation chain

full rationale

The paper reports an empirical evaluation: a fixed UNet is trained on five datasets at multiple input resolutions while holding other hyperparameters constant, then performance is measured with Dice and a width-stratified sensitivity derived from native-resolution distance transforms. No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the presented work. All reported effects (e.g., thin-vessel sensitivity drops of up to 15.8 pp) are direct experimental outcomes rather than quantities that reduce to the inputs by construction. The study is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The study rests on standard assumptions about UNet behavior under resolution change and on the accuracy of Euclidean-distance-based vessel width estimation; no new physical entities are postulated.

free parameters (1)

vessel width thresholds
Half-width <3 px for thin, 3-7 px for medium, >7 px for thick; chosen to stratify performance.

axioms (1)

domain assumption UNet segmentation performance depends primarily on input spatial resolution when other hyperparameters are held constant
Invoked to attribute observed sensitivity changes solely to downsampling.

pith-pipeline@v0.9.0 · 5586 in / 1180 out tokens · 40320 ms · 2026-05-13T20:44:37.103137+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

[1]

These alterations occur exclusively in the thinnest vessel branches, structures that are often only one to two pixels wide even at native acquisition resolution

Introduction The accurate detection of thin peripheral vessels in fundus images is critical for automated screening, as early microvascular alterations including capillary dropout, microaneurysms, and peripheral tortuosity changes are primary indicators of diabetic and hypertensiv e retinopathy [ 1, 2]. These alterations occur exclusively in the thinnest ...

work page
[2]

Galdran et al

Related Work Retinal vessel segmentation has been studied extensively across architectures, datasets, and loss functions [ 3, 7], however, input resolution as an independent variable has never been empirically examined. Galdran et al. [ 4] observed that a model trained on DRIVE failed to generalize to HRF and attributed this to the large resolution gap bu...

work page
[3]

Datasets Five publicly available fundus datasets were used in this study: DRIVE [ 11], STARE [14], CHASE_DB1 [15], HRF [16], and FIVES [ 17]

Datasets and Experimental Setup 3.1. Datasets Five publicly available fundus datasets were used in this study: DRIVE [ 11], STARE [14], CHASE_DB1 [15], HRF [16], and FIVES [ 17]. Their properties are summarized in Table 1. Native image widths span a 6x range, from 565 pixels to 3504 pixels, covering low resolution, mid resolution, and high resolution acqu...

work page
[4]

Resizing Conditions Five resizing conditions were defined per dataset

2048 x 2048 Not specified in dataset paper Adults: 200 DR, 200 AMD, 200 glaucoma, 200 normal, Zhejiang, China R1 to R5 3.6x 3.2. Resizing Conditions Five resizing conditions were defined per dataset. Table 2 lists all conditions and their corresponding processed sizes and downsampling ratios. Conditions were selected to cover the range from native resolut...

work page 2048
[5]

Results by Condition Table 3 reports the mean 5-fold results for all datasets and conditions

Results 4.1. Results by Condition Table 3 reports the mean 5-fold results for all datasets and conditions. The best thin vessel sensitivity per dataset is shown in bold. Across all five datasets and 25 conditions, thin vessel sensitivity ranges from 0.5139 to 0.6723, while Dice ranges from 0.5803 to 0.8394. The divergence between these two metrics across ...

work page
[6]

Resolution Selection Relative to Native Image Size The results reveal a resolution -dependent pattern in thin vessel sensitivity

Discussion 5.1. Resolution Selection Relative to Native Image Size The results reveal a resolution -dependent pattern in thin vessel sensitivity. For high -resolution datasets (HRF, FIVES), where native widths far exceed the encoder's receptive field capacity, downsampling consistently improves thin vessel detection. HRF th in vessel sensitivity improves ...

work page
[7]

Conclusion This paper presented a controlled empirical study of the effect of input resolution on retinal vessel segmentation performance across five publicly available fundus datasets spanning a 6x range of native image widths. A fixed UNet architecture (1.9M parame ters) was trained at five downsampling ratios per dataset with all other settings held id...

work page arXiv
[8]

Ophthalmology 128:1580-1591, 2021

Teo ZL, Tham YC, Yu M, Chee ML, Rim TH, Cheung N, Bikbov MM, Wang YX, Tang Y, Lu Y, Wong IY, Ting DSW, Tan GSW, Jonas JB, Sabanayagam C, Wong TY, Cheng CY: Global prevalence of diabetic retinopathy and projection of burden through 2045. Ophthalmology 128:1580-1591, 2021

work page 2045
[9]

Lancet 376:124-136, 2010

Cheung N, Mitchell P, Wong TY: Diabetic retinopathy. Lancet 376:124-136, 2010

work page 2010
[10]

Eng Appl Artif Intell 128:107454, 2024

Qing Q, Chen Y: A review of retinal vessel segmentation for fundus image analysis. Eng Appl Artif Intell 128:107454, 2024

work page 2024
[11]

Sci Rep 12:6174, 2022

Galdran A, Anjos A, Dolz J, Chakor H, Lombaert H, Ben Ayed I: State-of-the-art retinal vessel segmentation with minimalistic models. Sci Rep 12:6174, 2022

work page 2022
[12]

J Imaging Inform Med 38:520-533, 2024

Bhimavarapu U: Retina blood vessels segmentation and classification with the multi -featured approach. J Imaging Inform Med 38:520-533, 2024

work page 2024
[13]

J Imaging Inform Med 38:496-519, 2024

Cai P, Li B, Yan J: DEAF -Net: Detail -enhanced attention feature fusion network for retinal vessel segmentation. J Imaging Inform Med 38:496-519, 2024

work page 2024
[14]

IEEE Access, DOI: 10.1109/ACCESS.2024.3477420, 2024

Islam S, Deo RC, Barua PD, Soar J, Yu P: Retinal health screening using artificial intelligence with digital fundus images: A review of the last decade (2012 –2023). IEEE Access, DOI: 10.1109/ACCESS.2024.3477420, 2024

work page doi:10.1109/access.2024.3477420 2012
[15]

arXiv:2406.14994, 2024

Fadugba J, Köhler P, Koch L, Manescu P, Berens P: Benchmarking retinal blood vessel segmentation models for cross-dataset and cross-disease generalization. arXiv:2406.14994, 2024

work page arXiv 2024
[16]

Proc IEEE CVPR:936-944, 2017

Lin TY, Dollár P, Girshick R, He K, Hariharan B, Belongie S: Feature pyramid networks for object detection. Proc IEEE CVPR:936-944, 2017

work page 2017
[17]

Neural Netw 126:153-162, 2020

Wu Y, Xia Y, Song Y, Zhang Y, Cai W: NFN+: A novel network followed network for retinal vessel segmentation. Neural Netw 126:153-162, 2020

work page 2020
[18]

IEEE Trans Med Imaging 23:501-509, 2004

Staal J, Abramoff MD, Niemeijer M, Viergever MA, van Ginneken B: Ridge -based vessel segmentation in color images of the retina. IEEE Trans Med Imaging 23:501-509, 2004

work page 2004
[19]

Med Image Anal 70:102025, 2021

Wu H, Wang W, Zhong J, Lei B, Wen Z, Qin J: SCS -Net: A scale and context sensitive network for retinal vessel segmentation. Med Image Anal 70:102025, 2021

work page 2021
[20]

Sci Data 9:397, 2022

Lyu X, Cheng L, Zhang S: The RETA Benchmark for Retinal Vascular Tree Analysis. Sci Data 9:397, 2022

work page 2022
[21]

IEEE Trans Med Imaging 19:203-210, 2000

Hoover A, Kouznetsova V, Goldbaum M: Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Trans Med Imaging 19:203-210, 2000

work page 2000
[22]

Invest Ophthalmol Vis Sci 50:2004-2010, 2009

Owen CG, Rudnicka AR, Mullen R, Barman SA, Monekosso D, Whincup PH, Ng J, Paterson C: Measuring retinal vessel tortuosity in 10 -year-old children: validation of the CAIAR program. Invest Ophthalmol Vis Sci 50:2004-2010, 2009

work page 2004
[23]

Int J Biomed Imaging 2013:154860, 2013

Budai A, Bock R, Maier A, Hornegger J, Michelson G: Robust vessel segmentation in fundus images. Int J Biomed Imaging 2013:154860, 2013

work page 2013
[24]

Sci Data 9:475, 2022

Jin K, Huang X, Zhou J, Li Y, Yan Y, Sun Y, Zhang Q, Wang Y, Ye J: FIVES: A fundus image dataset for AI-based vessel segmentation. Sci Data 9:475, 2022

work page 2022
[25]

Lect Notes Comput Sci 9351:234-241, 2015

Ronneberger O, Fischer P, Brox T: U-Net: Convolutional networks for biomedical image segmentation. Lect Notes Comput Sci 9351:234-241, 2015

work page 2015
[26]

Proc ICLR, 2019

Loshchilov I, Hutter F: Decoupled weight decay regularization. Proc ICLR, 2019

work page 2019
[27]

Proc IEEE CVPR:16555-16564, 2021

Shit S, Paetzold JC, Sekuboyina A, Ezhov I, Unger A, Zhylka A, Pluim JPW, Bauer U, Menze BH: clDice: A novel topology-preserving loss function for tubular structure segmentation. Proc IEEE CVPR:16555-16564, 2021

work page 2021
[28]

Proc ICML:46640-46671, 2024

Stucki N, Paetzold JC, Shit S, Menze B, Bauer U: Topograph: An efficient graph-based framework for strictly topology preserving image segmentation. Proc ICML:46640-46671, 2024

work page 2024
[29]

Lect Notes Comput Sci 15008:721-731, 2024

Berger AH, Lux L, Stucki N, Bauer U, Menze B, Paetzold JC: Topologically faithful multi-class segmentation using Betti matching. Lect Notes Comput Sci 15008:721-731, 2024

work page 2024