Evidence-Guided Selection of Retinal Vessel Segmentation Pipelines: A Cross-Dataset Benchmark of Performance, Structure, Failure Modes, and Reproducibility

Ruoyu Yan; Yuejia Liu; Shixin Xu; Xiaoyan Li; Xi Luo; Yuhui Deng; Huaxiong Huang

cs.CVOct 22, 2026

Evidence-Guided Selection of Retinal Vessel Segmentation Pipelines: A Cross-Dataset Benchmark of Performance, Structure, Failure Modes, and Reproducibility

Authors: Ruoyu Yan, Yuejia Liu, Shixin Xu, Xiaoyan Li, Xi Luo, Yuhui Deng, Huaxiong Huang

Organizations: Duke Kunshan University, Kunshan, China · University of Toronto, Toronto, Canada · Beijing Normal–Hong Kong Baptist University, Zhuhai, China · York University, Toronto, Canada

Abstract

Retinal vessel segmentation is widely used in retinal image analysis, yet model selection based only on aggregate overlap scores can overlook structural fidelity, cross-dataset robustness, and application-specific error costs. We evaluated five executable binary vessel-segmentation pipelines (FSG-Net, SA-UNetv2, DCP-1024, GAVE, and FR-UNet) on 244 images from six public datasets using pixel-level agreement, cross-dataset stability, thin-vessel retention, structural metrics, error composition, and available computational evidence. Results were aggregated with equal weighting across datasets. No pipeline was optimal in every setting. FSG-Net achieved the highest equal-dataset mean Macro-F1 (0.77), IoU (0.63), clDice (0.83), and Skeleton F1 (0.78), whereas SA-UNetv2 achieved the highest mean sensitivity (0.85) and Thin-Vessel Sensitivity (0.63). Differences across datasets were often larger than those among the leading pipelines, and errors shared by all five methods were concentrated near annotated vessel boundaries and faint peripheral vessels. In an ancillary analysis on 45 HRF images, segmentation masks were converted into 11 ring-wise vascular features for leave-one-out classification. Controlled perturbations showed that local mask errors can be amplified in downstream features, while targeted stabilization reduced sensitivity for five features. The results support task-conditioned, benchmark-relative pipeline selection rather than a single universal ranking and highlight the need to validate the selected pipeline on the intended target domain.

Figures & tables

Figure 1.1: Representative HRF fundus images and pixel-level vessel segmentation results.

Dataset	$n$	Resolution	Source
AFIO	99	$1504\times 1000$	[ 1 ]
CHASEDB1	28	$999\times 960$	[ 5 ]
DualModal2019	30	$1024\times 1024$	[ 31 ]
HRF	45	$3504\times 2336$	[ 3 ]
LES	22	$1620\times 1444^{*}$	[ 24 ]
STARE	20	$700\times 605$	[ 11 ]

Table 2.1: Evaluation datasets used in the benchmark.

n

denotes the number of retained image–mask pairs. DualModal2019 was evaluated using the RGB image–mask pair.

Benchmark pipeline	Source method	Source publication	Public benchmark route
FSG-Net	FSG-Net	Full-scale Representation Guided Network for Retinal Vessel Segmentation	Official training and inference code
FR-UNet	FR-UNet + DTI 1	Full-Resolution Network and Dual-Threshold Iteration for Retinal Vessel and Coronary Angiograph Segmentation	Official training and inference code
SA-UNetv2	SA-UNetv2	SA-UNetv2: Rethinking Spatial Attention U-Net for Retinal Vessel Segmentation	Official training and inference code
DCP-1024	DCP	Convolutional Prompting for Broad-Domain Retinal Vessel Segmentation	Released 1024-resolution checkpoint and inference route
GAVE	RIP-AV	RIP-AV: Joint Representative Instance Pre-training with Context-Aware Network for Retinal Artery/Vein Segmentation	Released checkpoint and inference code

Table 2.2: Included pipelines, source provenance, and reproducible benchmark routes.

Pipeline	Evidence source	Local setup	Model / compute	Best val. ckpt	Memory / training time
FR-UNet	FIVES local training + profiling	$512\times 512$ ; BS=4; AMP	7.38 M params; 430.77 GFLOPs/img; 28.32 MiB ckpt	Epoch 47; val. F1=0.93	Peak VRAM: 7.08 GB; median 44.57 s/epoch; best ckpt: 35.10 min; total: 52.16 min
SA-UNetv2	FIVES local log + profiling	$1028\times 1028$ ; BS=2	0.26 M params; 78.18 GFLOPs/img; 3.24 MiB ckpt	Epoch 55; val. F1=0.93	Peak VRAM: 13.18 GB; mean 13.45 s/epoch (75 epochs) 2 ; best-ckpt time: N/A; total: 16.85 min
FSG-Net	FIVES local training + profiling	$288\times 288$ ; BS=4; FP32	16.85 M params; 199.61 GFLOPs/img; 64.45 MiB ckpt	Epoch 181; val. F1=0.90	Peak VRAM: 7.94 GB; median 34.91 s/epoch; best ckpt: 105.37 min; total: 116.42 min
DCP-1024	Released ckpt + published complexity	$1024\times 1024$ local inference; no local training	12.50 M params 1 ; FLOPs/ckpt not locally profiled	N/A	Training VRAM/time: N/A
GAVE	Released ckpt	Overlapping $256\times 256$ patch inference; no local training	Params/FLOPs/ckpt not locally profiled	N/A	Training VRAM/time: N/A

Table 3.1: Available computational-cost and local-training evidence for the evaluated pipelines.

Metric	What it measures	Interpretation
Accuracy	Fraction of correctly classified pixels.	Can be misleading because retinal vessel segmentation is strongly class-imbalanced.
F1 / $F_{\beta}$	Balance between vessel precision and sensitivity.	F1 gives equal weight to precision and sensitivity; $F_{\beta}$ allows different FN–FP priorities. Image-level F1 is the standard pixel-level F1 computed separately for each image, and Macro-F1 summarizes these image-level values under the benchmark aggregation procedure.
Sensitivity	Fraction of reference vessel pixels recovered.	Higher values mean fewer false negatives.
Precision	Fraction of predicted vessel pixels that are correct.	Higher values mean fewer false positives.
IoU	Overlap between predicted and reference vessel regions.	Higher values indicate stronger region-level agreement.
MCC	Agreement considering all four confusion-matrix components.	Useful under class imbalance because true positives, true negatives, false positives, and false negatives all contribute.

Table 4.1: Overview of the principal evaluation metrics used in the benchmark.

Figure 4.1: Dataset-level vessel occupancy and the apparent Accuracy of an all-background prediction.

Figure 4.2: Nearly matched pixel-level F1 values with divergent structural fidelity and spatial error profiles. White denotes the supplied reference vessel, black denotes false-negative (FN) pixels, and blue denotes false-positive (FP) pixels.

Evidence dimension	Measures	Role in model selection
Pixel-level performance	F1/ $F_{\beta}$ , sensitivity, precision, IoU, and MCC	Characterizes overall pixel-wise agreement; supports selection for overall performance, low-FN, low-FP, or balanced objectives.
Cross-dataset robustness	Equal-dataset mean $F_{\beta}$ , cross-dataset standard deviation, and worst-dataset performance	Quantifies expected performance, cross-dataset variability, and downside risk under domain uncertainty.
Thin-vessel retention	Thin-Vessel Sensitivity (TVS)	Measures recovery of vessels and supports selection when preserving thin vessels or reducing missed branches is a priority.
Structural fidelity	clDice and tolerant Skeleton F1 (SF1)	Assesses centerline containment and tolerant skeleton alignment; supports selection when focusing on vascular continuity and structural fidelity.

Table 4.2: Four evidence dimensions used for task-conditioned model selection.

Scenario	$\bm{\beta}$	Weights (P/T/S/R)	$\bm{\rho}$	First / second choice	Evidence gap	Decision
Mixed-objective default	1	50/15/20/15	0.30	SA-UNetv2 / FSG-Net	75.82 vs. 66.76 ( $+9.06$ )	Provisional
Pixel-level mean	1	100/0/0/0	0.00	FSG-Net / DCP-1024	77.09 vs. 76.40 ( $+0.69$ )	Two-model shortlist
Low FN / thin vessels	2	50/50/0/0	0.30	SA-UNetv2 / FSG-Net	68.01 vs. 66.99 ( $+1.02$ )	SA-UNetv2 preferred
Structural fidelity	1	0/0/100/0	0.30	FSG-Net / SA-UNetv2	79.46 vs. 78.85 ( $+0.61$ )	Two-model shortlist
Unknown-domain robustness	1	100/0/0/0	0.70	SA-UNetv2 / FSG-Net	71.95 vs. 71.81 ( $+0.15$ )	Two-model shortlist
Resource-oriented	1	30/0/0/70	0.30	SA-UNetv2 / insufficient evidence	86.94; no valid second-place comparison	Provisional

Table 5.1: Preset decision rules implemented in the public model-selection tool. Weights are reported as pixel/thin-vessel/structure/resource percentages.

Figure 5.1: Public browser-based decision tool for task-conditioned retinal vessel segmentation pipeline selection. Each framework component in panel A corresponds by row to the interface excerpt in panel B: versioned benchmark evidence (B1), custom decision controls (B2), task-conditioned ranking output (B3), and post-ranking residual-risk profiles (B4). The tool recomputes scenario-specific rankings over fixed benchmark evidence; it does not imply a universal pipeline ranking.

Model	Macro-F1	Sensitivity	Precision	MCC	IoU
DCP-1024	0.76	0.79	0.77	0.76	0.62
FR-UNet	0.73	0.77	0.72	0.72	0.58
FSG-Net	0.77	0.84	0.73	0.76	0.63
GAVE	0.73	0.80	0.70	0.72	0.58
SA-UNetv2	0.76	0.85	0.71	0.75	0.62

Table 5.2: Mean pixel-level segmentation performance across the evaluation datasets. Bold values indicate the best performance for each metric.

Figure 5.2: Dataset-level Macro-F1 across five evaluated pipelines and six evaluation datasets. Bold values indicate the highest Macro-F1 within each dataset.

Figure 5.3: Distribution of within-dataset image-level F1 variability across the six evaluation datasets.

Model	Equal-dataset Macro-F1	Cross-dataset SD	Worst Macro-F1	Worst dataset
FSG-Net	0.77	0.05	0.70	AFIO
DCP-1024	0.76	0.07	0.64	AFIO
SA-UNetv2	0.76	0.05	0.70	AFIO
GAVE	0.73	0.06	0.65	AFIO
FR-UNet	0.73	0.05	0.66	AFIO

Table 5.3: Cross-dataset robustness of the evaluated pipelines. Bold values indicate the best result for each robustness measure.

Model	Mean TVS	TVS SD	Worst-case TVS (dataset)	Mean Sensitivity
SA-UNetv2	0.63	0.12	0.52 (HRF)	0.85
FSG-Net	0.62	0.12	0.51 (AFIO)	0.84
DCP-1024	0.52	0.13	0.40 (HRF)	0.79
GAVE	0.51	0.13	0.35 (CHASEDB1)	0.80
FR-UNet	0.47	0.18	0.29 (LES)	0.77

Table 5.4: Equal-dataset thin-vessel retention summary. Means and standard deviations were calculated from the six dataset-level values with equal dataset weighting. TVS variability should be interpreted jointly with the mean and worst-case TVS. Bold values indicate the strongest selection-relevant results.

Figure 5.4: Thin-Vessel Sensitivity across evaluated pipelines and datasets.

Figure 5.5: Cross-dataset structural fidelity measured by clDice (left) and tolerant Skeleton F1 with a 2-pixel tolerance (right).

Model	Mean clDice $\pm$ SD	Worst-case clDice (dataset)	Mean SF1 $\pm$ SD	Worst-case SF1 (dataset)
FSG-Net	0.83 $\pm$ 0.04	0.77 (AFIO)	0.78 $\pm$ 0.11	0.63 (HRF)
SA-UNetv2	0.82 $\pm$ 0.03	0.77 (AFIO)	0.76 $\pm$ 0.13	0.57 (HRF)
DCP-1024	0.81 $\pm$ 0.06	0.70 (AFIO)	0.75 $\pm$ 0.15	0.54 (HRF)
FR-UNet	0.77 $\pm$ 0.06	0.69 (HRF)	0.69 $\pm$ 0.21	0.35 (HRF)
GAVE	0.76 $\pm$ 0.05	0.68 (AFIO)	0.74 $\pm$ 0.08	0.61 (AFIO)

Table 5.5: Equal-dataset centerline and structural-fidelity summary. Means and standard deviations were calculated from the six dataset-level values with equal dataset weighting. clDice and SF1 represent complementary structural criteria and are not interpreted as interchangeable measures. Bold values indicate the highest mean and worst-case structural scores; SDs are reported as measures of cross-dataset variability.

Error type	Operational definition	Model-selection interpretation
Associated metric signal	FN-oriented errors are reflected by Sensitivity and TVS, FP-oriented errors by Precision, and centerline or continuity errors by clDice and SF1	Links each error type to the metric dimensions used in the benchmark; the taxonomy is complementary to, not a replacement for, the aggregate metrics.
Boundary shift	Paired FN and FP pixels with one-pixel 8-neighbor adjacency; the shared count is defined as $\min(\text{adjacent FN},\,\text{adjacent FP})$	Represents a localized boundary displacement and avoids double-counting the same spatial mismatch as independent FN and FP errors.
Too thin	Residual adjacent FN pixels remaining after shared boundary-shift pairing	Indicates localized underestimation of vessel width.
Missed vessel	FN pixels with no predicted-mask pixel in their 8-neighborhood	Indicates vessel omission under the documented legacy classification rule.
Too thick	Residual adjacent FP pixels remaining after shared boundary-shift pairing	Indicates localized overestimation of vessel width.
Extra vessel	FP pixels with no ground-truth-mask pixel in their 8-neighborhood	Indicates an unsupported vessel response under the documented legacy classification rule.

Table 5.6: Structural error taxonomy and its interpretation for model selection.

Figure 5.6: Representative FSG-Net error patterns on HRF under the structural error taxonomy. Black denotes false-negative (FN) pixels, and blue denotes false-positive (FP) pixels. The original fundus image is retained as the background.

Figure 5.7: Model-wise composition of residual structural-error pixels. The annotation row indicates the evaluation metrics most directly associated with each error category; these associations are complementary rather than one-to-one.

Dataset	Thin-skeleton share	Vessel occupancy (visible FOV)	Mean thin-vessel diameter	Mean thick-vessel diameter
AFIO	0.90%	0.15	2.43	7.80
CHASEDB1	30.60%	0.10	2.18	7.77
DualModal2019	39.00%	0.08	2.17	6.48
HRF	50.10%	0.09	4.16	13.35
LES	25.30%	0.10	4.13	11.17
STARE	22.00%	0.08	2.00	4.84

Table 6.1: Benchmark-derived dataset morphology descriptors. All morphology descriptors were recomputed from the exact reference masks used in the benchmark.

Figure 6.1: Model-centered relative Macro-F1 across datasets.

Dataset	Mean TVS $\pm$ across-model SD	Mean Sensitivity $\pm$ across-model SD
AFIO	0.48 $\pm$ 0.05	0.57 $\pm$ 0.05
CHASEDB1	0.50 $\pm$ 0.11	0.85 $\pm$ 0.03
DualModal2019	0.80 $\pm$ 0.06	0.95 $\pm$ 0.01
HRF	0.45 $\pm$ 0.08	0.81 $\pm$ 0.04
LES	0.52 $\pm$ 0.14	0.85 $\pm$ 0.06
STARE	0.54 $\pm$ 0.09	0.83 $\pm$ 0.05

Table 6.2: Dataset-level thin-vessel recoverability. TVS and sensitivity values are reported as the across-model mean

\pm

standard deviation.

Dataset	Mean clDice $\pm$ across-model SD	Mean SF1( $r=2$ ) $\pm$ across-model SD
AFIO	0.73 $\pm$ 0.04	0.63 $\pm$ 0.03
CHASEDB1	0.81 $\pm$ 0.04	0.79 $\pm$ 0.04
DualModal2019	0.82 $\pm$ 0.02	0.86 $\pm$ 0.02
HRF	0.76 $\pm$ 0.04	0.56 $\pm$ 0.14
LES	0.84 $\pm$ 0.03	0.74 $\pm$ 0.05
STARE	0.83 $\pm$ 0.04	0.88 $\pm$ 0.05

Table 6.3: Dataset-level centerline and skeleton fidelity. clDice and SF1(

r=2

) are reported as across-model mean

\pm

standard deviation.

Figure 6.2: Dataset- and model-conditioned composition of structural error types.

Figure 7.1: Complete analysis pipeline from segmentation mask to vascular features and downstream classification. The same original image is processed with each segmentation mask separately; the classifier receives only the ring-wise vascular features generated from the corresponding mask.

Method	Correct / total	Errors	Accuracy	95% Wilson CI
GAVE	34/45	11	0.76	[0.61, 0.86]
GT	33/45	12	0.73	[0.59, 0.84]
SA-UNetv2	29/45	16	0.64	[0.50, 0.77]
FR-UNet	28/45	17	0.62	[0.48, 0.75]
FSG-Net	27/45	18	0.60	[0.45, 0.73]
DCP-1024	24/45	21	0.53	[0.39, 0.67]

Table 7.1: Baseline vascular-only classification performance.

Perturbation	Baseline train $\rightarrow$ perturbed test	Perturbed LOOCV	Interpretation
No perturbation (GT)	0.73	–	Original vascular-feature baseline.
mask_dilate1	0.42	0.71	Global shifts in vessel area and caliber substantially reduce performance when the original classifier is applied without re-adaptation.
mask_erode1	0.44	0.87	Erosion alters thin-vessel representation and connectivity, while substantial class-discriminative information remains after re-adaptation.
skeleton_prune_tips5	0.56	0.80	Removal of terminal branches changes geometric and topological features, while much of the classification signal remains recoverable after re-adaptation.

Table 7.2: Summary of vascular-feature classification under controlled mask perturbations.

Feature	Stabilization procedure	Original sensitivity	Processed sensitivity	Relative reduction	Interpretation
Calibre diameter IQR	Spline centerline + Gaussian-smoothed EDT	0.3528	0.2897	17.9%	Sensitivity decreased, indicating a clear improvement in stability.
Tortuosity IQR	Spline centerline	2.9719	0.9037	69.6%	Largest sensitivity reduction among the five evaluated features.
Median signed curvature	Spline centerline	3.0228	1.2569	58.4%	Substantially reduced sensitivity to local directional jitter.
Median absolute curvature	Spline centerline	2.6024	1.2569	51.7%	Reduced sensitivity to pixel-level centerline jaggedness.
VCC	Graph pruning	0.3107	0.2911	6.3%	Modest stabilization, primarily through removal of short terminal spurs.

Table 7.3: Sensitivity of selected vascular features before and after targeted stabilization. Relative reduction is computed with respect to the original sensitivity; larger reductions indicate greater stabilization.

Method	Correct / total	Errors	Accuracy	95% Wilson CI
GT	35/45	10	0.78	[0.64, 0.87]
FSG-Net	30/45	15	0.67	[0.52, 0.79]
SA-UNetv2	29/45	16	0.64	[0.50, 0.77]
FR-UNet	27/45	18	0.60	[0.45, 0.73]
DCP-1024	26/45	19	0.58	[0.43, 0.71]
GAVE	25/45	20	0.56	[0.41, 0.69]

Table 7.4: Vascular-only classification performance after selective feature stabilization.

Figure 8.1: Proportion of shared false-negative and false-positive errors located within normalized ground-truth boundary bands as a function of cross-model consensus level.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Pipeline(s)	Relevant training / exposure evidence	Dataset handling	Interpretive qualification
FSG-Net; FR-UNet; SA-UNetv2	Locally supervised training on FIVES was performed for these benchmark realizations.	FIVES excluded	Direct overlap between local training and benchmark evaluation was avoided.
DCP-1024	FIVES was included in the source broad-domain development setting; this benchmark used the released checkpoint rather than local retraining.	FIVES excluded	Source-development exposure is retained as a qualification and is not treated as absent.
GAVE / RIP-AV vessel channel	The released checkpoint used for GAVE inference in this benchmark was trained on DRIVE.	DRIVE excluded	The reported GAVE results are qualified by checkpoint exposure and should not be interpreted as an external test on DRIVE.

Appendix

Table A.1: Training-exposure ledger used to determine evaluation-dataset inclusion. Source-paper exposure and local checkpoint provenance are distinguished so that dataset exclusions are not overstated as a controlled common-training comparison.

Figure B.1: Sensitivity of equal-dataset Macro-F1 to a common single-threshold operating point.

Pipeline	Tested factors	Rationale
SA-UNetv2	DropBlock rate; scheduler patience	Tests regularization strength and the timing of learning-rate reduction.
FSG-Net	Weight decay; crop size	Tests regularization strength and the effect of spatial training context.
FR-UNet	Dropout rate; batch size	Tests regularization strength and batch-dependent optimization behavior.

Appendix

Table C.1: Tested configuration factors and their rationale. DCP-1024 and GAVE were not included in additional configuration sweeps because reproducible local training routes were unavailable.

Configuration	AFIO	CHASEDB1	DualModal2019	HRF	LES	STARE	Equal-dataset Macro-F1
DropBlock $=0.05$	0.6827	0.7906	0.7413	0.7272	0.8234	0.8068	0.7620
DropBlock $=0.10$	0.6778	0.7824	0.7369	0.7224	0.8249	0.8069	0.7586
Patience $=6$	0.6917	0.7911	0.7387	0.7257	0.8224	0.8089	0.7631
Patience $=14$	0.6841	0.7890	0.7412	0.7273	0.8223	0.8048	0.7614

Appendix

Table C.2: SA-UNetv2 configuration sensitivity across the six evaluation datasets. Fixed settings included a

1028\times 1028

input size, batch size 2,

0.5\,\mathrm{BCE}+0.5\,\mathrm{MCC}

loss, DropBlock block size 7, no data augmentation, and split seed 23.

Configuration	AFIO	CHASEDB1	DualModal2019	HRF	LES	STARE	Equal-dataset Macro-F1
WD $=0.01$	0.6984	0.7787	0.7252	0.8002	0.8239	0.8012	0.7713
WD $=0.10$	0.6899	0.7836	0.7317	0.8025	0.8309	0.8081	0.7744
Crop $=224\times 224$	0.6925	0.7694	0.7270	0.7932	0.8175	0.7867	0.7644
Crop $=352\times 352$	0.6967	0.7906	0.7302	0.8032	0.8295	0.8152	0.7776

Appendix

Table C.3: FSG-Net configuration sensitivity across the six evaluation datasets. The source-aligned setting used weight decay 0.05 and

288\times 288

random crops.

Configuration	AFIO	CHASEDB1	DualModal2019	HRF	LES	STARE	Equal-dataset Macro-F1
Dropout $=0.10$	0.6944	0.7869	0.7351	0.7393	0.8197	0.8126	0.7646
Dropout $=0.30$	0.7014	0.7674	0.7017	0.7066	0.8091	0.7937	0.7466
Batch size $=2$	0.6818	0.7885	0.7185	0.7291	0.8190	0.7863	0.7539

Appendix

Table C.4: FR-UNet configuration sensitivity across the six evaluation datasets. The source-aligned setting used a dropout rate of 0.20 and batch size 4.

Metric	Definition	Benchmark use
Sensitivity	$\displaystyle\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$	Vessel recall; FN-oriented evidence.
Precision	$\displaystyle\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$	Prediction reliability; FP-oriented evidence.
F1	$\displaystyle\frac{2\,\mathrm{TP}}{2\,\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$	Primary pixel-level agreement measure.
$F_{\beta}$	$\displaystyle F_{\beta}=(1+\beta^{2})\frac{PR}{\beta^{2}P+R}$	Scenario-dependent precision–recall weighting; $\beta>1$ places greater emphasis on recall and therefore on reducing FN errors.
IoU	$\displaystyle\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$	Overlap-based consistency check.
MCC	$\displaystyle\frac{\mathrm{TP}\,\mathrm{TN}-\mathrm{FP}\,\mathrm{FN}}{\sqrt{(\mathrm{TP}+\mathrm{FP})(\mathrm{TP}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})}}$	Class-imbalance-aware consistency check.

Appendix

Table D.1: Pixel-level metric definitions and their use in the benchmark.

Metric	Definition	Benchmark setting and interpretation
Thin-Vessel Sensitivity (TVS)	$\begin{aligned} D_{G}(x)&=2\,DT_{G}(x),\\ \tau_{\mathrm{thin}}&=\max\!\left(1\,\mathrm{px},\,0.018D_{\mathrm{OD}}\right),\\ T_{G}^{\mathrm{thin}}&=\left\{x\in S_{G}:D_{G}(x)\leq\tau_{\mathrm{thin}}\right\},\\ \mathrm{TVS}&=\frac{\left\|T_{G}^{\mathrm{thin}}\cap P\right\|}{\left\|T_{G}^{\mathrm{thin}}\right\|}.\end{aligned}$	GT-conditioned recovery of fine vessels defined on the ground-truth skeleton. The diameter threshold is normalized by the optic-disc diameter $D_{\mathrm{OD}}$ , with a minimum threshold of $1$ pixel.
clDice	$\begin{aligned} T_{\mathrm{prec}}&=\frac{\|S_{P}\cap G\|}{\|S_{P}\|},\\ T_{\mathrm{sens}}&=\frac{\|S_{G}\cap P\|}{\|S_{G}\|},\\ \mathrm{clDice}&=\frac{2T_{\mathrm{prec}}T_{\mathrm{sens}}}{T_{\mathrm{prec}}+T_{\mathrm{sens}}}.\end{aligned}$	Bidirectional skeleton-to-region containment, following the clDice formulation [ 28 ] .
Tolerant Skeleton F1 (SF1)	$\begin{aligned} P_{r}&=\frac{\left\|\left\{p\in S_{P}:d(p,S_{G})\leq r\right\}\right\|}{\|S_{P}\|},\\ R_{r}&=\frac{\left\|\left\{g\in S_{G}:d(g,S_{P})\leq r\right\}\right\|}{\|S_{G}\|},\\ \mathrm{SF1}_{r}&=\frac{2P_{r}R_{r}}{P_{r}+R_{r}}.\end{aligned}$	Symmetric tolerance-based skeleton alignment. The main analysis used $r=2\,\mathrm{px}$ .

Appendix

Table D.2: Thin-vessel and structural-fidelity metric definitions and benchmark settings.

Summary	Definition	Interpretation
Dataset-level mean	$\displaystyle\bar{x}_{md}=\frac{1}{n_{d}}\sum_{i=1}^{n_{d}}x_{mdi}$	Average across images within a given model–dataset pair.
Equal-dataset cross-dataset mean	$\displaystyle\bar{x}_{m}=\frac{1}{D}\sum_{d=1}^{D}\bar{x}_{md}$	Assigns equal weight to each dataset regardless of its number of images.
Within-dataset image-level SD	$\displaystyle s_{md}=\sqrt{\frac{1}{n_{d}-1}\sum_{i=1}^{n_{d}}\left(x_{mdi}-\bar{x}_{md}\right)^{2}}$	Measures image-level variability within a given model–dataset pair.
Cross-dataset SD	$\displaystyle s_{m}=\sqrt{\frac{1}{D-1}\sum_{d=1}^{D}\left(\bar{x}_{md}-\bar{x}_{m}\right)^{2}}$	Measures variability across the dataset-level means.
Worst-dataset result	$\displaystyle x_{m}^{\mathrm{worst}}=\min_{d}\bar{x}_{md}$	Represents the lowest observed dataset-level performance for model $m$ .

Appendix

Table D.3: Image-level, dataset-level, and cross-dataset aggregation rules.

Figure E.1: Sensitivity of structural-fidelity scores to the structural weight $\eta$ . All other structural-preset settings were held fixed, including $\rho=0.30$ , while $\eta$ varied from 0 to 1. The vertical line indicates the prespecified value $\eta=0.80$ .

$\bm{\rho}$	Mixed-objective	Pixel performance	Low-FN / thin vessels	Structural fidelity	Resource-oriented
0.0	SA-UNetv2 / FSG-Net	FSG-Net / DCP-1024	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE
0.1	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE
0.2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE
0.3	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE
0.4	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE
0.5	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / FSG-Net	FSG-Net / SA-UNetv2	SA-UNetv2 / GAVE

Appendix

Table E.1: Sensitivity of the top-two model recommendations to the cross-dataset risk weight

\rho

. Each cell reports the first- and second-ranked models (Top 1 / Top 2).

Figure F.1: Sensitivity of equal-dataset Macro-F1 rankings to GT-boundary-band exclusion.

Scope	Comparison baseline	Baseline F1	Vote F1	Mean $\Delta$ F1	Baseline IoU	Vote IoU	Error change
All datasets (244 images)	Per-image best single pipeline	0.76	0.83	$+0.08$	0.61	0.73	$-10.74$ M

Appendix

Table G.1: Aggregate performance of GT-informed oracle vote fusion across all 244 images. The baseline uses the best-performing single pipeline selected independently for each image.

Figure G.1: Oracle vote versus the best single model on HRF/15_dr.

Dataset	Images	$\Delta$ F1	$\Delta$ IoU	FN recovered	FP removed
AFIO	99	$+0.13$	$+0.17$	37.4%	51.0%
CHASEDB1	28	$+0.10$	$+0.14$	55.6%	44.4%
DualModal2019	30	$+0.10$	$+0.14$	54.4%	44.7%
HRF	45	$+0.11$	$+0.16$	48.7%	52.3%
LES	22	$+0.09$	$+0.13$	50.3%	51.2%
STARE	20	$+0.08$	$+0.12$	45.7%	45.8%

Appendix

Table H.1: Dataset-level performance of GT-informed Laplacian oracle repair.

Item	Recorded value
Loss under test	$0.4\,\mathrm{BCE}+0.4\,\mathrm{MCC}+0.2\,\mathrm{soft\text{-}clDice}$
Run/configuration	$1028\times 1028$ input; batch size 2; split seed 23
Evaluation set	HRF; 45 matched images
Pooled F1	0.7126115599
Mean image-level F1	0.7083935962
Interpretation	Retained as a negative feasibility result for modifying the loss function under this setup; clDice remains the core structural evaluation metric.

Appendix

Table I.1: HRF feasibility test for incorporating soft-clDice into the training loss.

Figure J.1: Ring-wise mask mechanism for VAF. Green denotes vessel pixels shared by GT and FSG, cyan denotes false negatives, and red denotes false positives. VAF is the number of vessel pixels inside the ring divided by the valid ring area.

Figure J.2: Pointwise interpretation of graph_edge_count and the skeleton graph. The red markers identify junction nodes, and graph edges link neighboring skeleton pixels in the 8-neighborhood. The edge count is the total number of such links within the selected ring.

Figure J.3: Diameter-sampling mechanism for caliber. EDT-derived diameters are sampled along the GT and FSG centerlines. caliber_diam_mm_IQR is the range of the middle 50% of these diameter samples.

Figure J.4: Pixel skeleton versus spline-smoothed centerline for VLD. The pixel polyline is replaced by a continuous arc-length-fitted spline within the 0.5–1.0 DD ring.

Figure J.5: The left panel shows diameters sampled from the Gaussian-smoothed EDT along the spline centerline. The right panel shows the raw EDT, smoothed EDT, and their difference. Both panels use the original binary mask with the mask boundary.

Figure J.6: Graph pruning: original skeleton graph versus terminal-spur pruning. The left panel shows the original graph; the right panel removes terminal spurs no longer than 5 pixels. Red nodes indicate local graph nodes, while the main trunk connectivity is preserved as far as possible.

Method	Top five feature types (mean absolute coefficient)
GT	1. VAF (0.301); 2. VCC (0.234); 3. $D_{\mathrm{ratio}}$ (0.221); 4. $L_{\mathrm{thin}}$ (0.206); 5. median absolute curvature (0.149)
FSG-Net	1. caliber diameter IQR (0.289); 2. median absolute curvature (0.231); 3. $L_{\mathrm{thin}}$ (0.201); 4. $VC_{\mathrm{mean}}$ (0.201); 5. $D_{\mathrm{ratio}}$ (0.189)
DCP-1024	1. caliber diameter IQR (0.296); 2. $D_{\mathrm{ratio}}$ (0.249); 3. $VC_{\mathrm{mean}}$ (0.229); 4. VAF (0.201); 5. median signed curvature (0.199)
FR-UNet	1. caliber diameter IQR (0.299); 2. $VC_{\mathrm{mean}}$ (0.276); 3. VCC (0.239); 4. VAF (0.218); 5. median absolute curvature (0.206)
GAVE	1. caliber diameter IQR (0.260); 2. $VC_{\mathrm{mean}}$ (0.236); 3. median absolute curvature (0.213); 4. $L_{\mathrm{thin}}$ (0.203); 5. median signed curvature (0.202)
SA-UNetv2	1. caliber diameter IQR (0.277); 2. $L_{\mathrm{thin}}$ (0.209); 3. $D_{\mathrm{ratio}}$ (0.201); 4. $VC_{\mathrm{mean}}$ (0.194); 5. VAF (0.194)

Appendix

Table J.1: Method-specific top five vascular feature types ranked by mean absolute coefficient magnitude, averaged across rings.

Retinal Vessel Segmentation

Retinal vessel segmentation identifies blood vessels in retinal images to support vascular measurement, disease analysis, and screening.

See all 24 papers

Mar 23, 2026

HMS-VesselNet: Hierarchical Multi-Scale Attention Network with Topology-Preserving Loss for Retinal Vessel Segmentation

Retinal vessel segmentation methods based on standard overlap losses tend to miss thin peripheral vessels because these structures occupy very few pixels and have low contrast against the background. We propose HMS-VesselNet, a hierarchical multi-scale network that processes fundus images across four parallel branches at different resolutions and combines their outputs using learned fusion weights. The training loss combines Dice, binary cross-entropy, and centerline Dice to jointly optimize area overlap and vessel continuity. Hard example mining is applied from epoch 20 onward to concentrate gradient updates on the most difficult training images. Tested on 68 images from DRIVE, STARE, and CHASE_DB1 using 5-fold cross-validation, the model achieves a mean Dice of 88.72 +/- 0.67%, Sensitivity of 90.78 +/- 1.42%, and AUC of 98.25 +/- 0.21%. In leave-one-dataset-out experiments, AUC remains above 95% on each unseen dataset. The largest improvement is in the recall of thin peripheral vessels, which are the structures most frequently missed by standard methods and most critical for early detection of diabetic retinopathy.

Amarnath R

United Institute of Technology, Coimbatore, India

Retinal Vessel Segmentation Multi-Scale Feature Fusion Topology-Preserving Image Segmentation

Ask PDF

Oct 15, 2026

FPG-Seg: Annotation-free Retinal Vessel Topology Segmentation

State-of-the art approaches to retinal vessel segmentation typically require extensive annotations, a highly time-consuming process, reliant on clinical expertise. We present FPG-Seg, an annotation-free pipeline for retinal vessel segmentation combining classical vesselness filtering, Bayesian deep learning and topology-aware graph reconstruction to segment retinal vessel topology without manual labels. Tested across four widely used benchmark datasets for retinal vessel segmentation, FPG-Seg attains a clDice of

0.74{-}0.78

, representing

86{-}98\%

of the performance of an equivalent fully-supervised baseline, despite using no manual annotations for training or hyperparameter tuning. As part of this work, we also provide, to our knowledge, the first performance benchmark of an annotation-free vessel segmentation method on FIVES.

James Willoughby, Irina Voiculescu

University of Oxford, Department of Computer Science

Retinal Vessel Segmentation Unsupervised Image Segmentation Topology-Preserving Image Segmentation

Ask PDF

Sep 22, 2026

Observer Choice and Threshold Selection in Retinal Vessel Segmentation: A Subject-Separated Evaluation

The annotation used to select a segmentation threshold is part of the evaluation protocol, yet its effect is easily conflated with model quality. We examine this choice for retinal vessel segmentation using all 28 CHASE DB1 images and both human annotations. A fixed seven-fold protocol keeps both eyes of each of the 14 subjects together. Random forests and Extra Trees are fitted against observer 1 with three random seeds, yielding 42 fits. Five threshold policies share identical score maps: fixed 0.50, observer-1 tuning, observer-2 tuning, mean-observer tuning, and maximin tuning of the per-image lower observer Dice. For random forests, maximin changes the threshold in 19 of 21 fits, but worst-observer Dice decreases from 70.53 percent to 70.45 percent. The paired difference is -0.073 percentage points, with a conditional subject-bootstrap 95 percent interval of [-0.384, 0.238]. Extra Trees shows the same direction. Identical observer-1-tuned random-forest masks score 73.66 percent against observer 1 and 71.06 percent against observer 2. The results support explicit reporting of both the threshold-selection reference and evaluation reference; they do not support an accuracy benefit from maximin tuning in this cohort. All splits, raw predictions, metrics and code are supplied. AI assistance is disclosed.

Wenhao Xu, Yixian Kong, Ting Pan +2

Zhengzhou Police University, Zhengzhou, China · School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China · Qilu University of Technology and Jinan Supercomputing Center, Jinan, China

Retinal Vessel Segmentation

Ask PDF

Abstract

Retinal Vessel Segmentation

Explore similar work

HMS-VesselNet: Hierarchical Multi-Scale Attention Network with Topology-Preserving Loss for Retinal Vessel Segmentation

FPG-Seg: Annotation-free Retinal Vessel Topology Segmentation

Observer Choice and Threshold Selection in Retinal Vessel Segmentation: A Subject-Separated Evaluation