PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.35357v1cs.CV

TL;DR

Retinal vessel segmentation model selection can miss structural fidelity, cross-dataset robustness, and application-specific error costs when based only on aggregate overlap scores. This study benchmarks executable pipelines across datasets and evaluation dimensions. Its findings support benchmark-relative, task-conditioned selection rather than a universal ranking.

  • Problem

    Aggregate overlap scores alone can overlook structural fidelity, cross-dataset robustness, and application-specific error costs in retinal vessel segmentation.

    Source: §1 · p. 2, §4.1 · p. 10

  • Method

    The study evaluates executable segmentation pipelines using pixel-level, thin-vessel, structural, cross-dataset, and resource evidence to support application-specific selection.

    Source: §1 · p. 3

  • Results

    Across predefined scenarios, FSG-Net and SA-UNetv2 accounted for all first-ranked recommendations, with preferences varying by the emphasized objective.

    Source: §8 · p. 29

  • Takeaways & Limitations

    The final pipeline choice should reflect dominant error costs, structural requirements, computational constraints, and the target domain.

    Source: §8 · p. 32

    The benchmark compares complete executable pipelines under differing input, preprocessing, inference, post-processing, and training conditions rather than architectures under strictly controlled training conditions.

    Source: §8 · p. 29

Abstract

from arXiv · show

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.

1 Introduction

This study frames retinal vessel segmentation as task-conditioned pipeline selection because pipelines can trade off overall agreement, thin-vessel recovery, and structural preservation. It evaluates executable pipelines across multiple evidence dimensions and examines how mask differences affect downstream vascular features.

Background and Model-Selection Context

  • Vessel visibility varies with resolution, contrast, illumination, acquisition conditions, and pathology, producing missed vessels, false positives, boundary shifts, and width errors.Different tasks may prioritize overall agreement, thin-vessel recovery, or structural preservation, so a pipeline preferred under one objective may not suit another.Source: §1 · p. 2

Existing Evaluation and Review Gaps and Motivation

  • Prior studies differ in datasets, preprocessing, thresholds, implementations, protocols, and reported metrics, limiting direct comparisons and practical selection guidance.Cross-dataset studies also report performance degradation under domain shift.Source: §1 · p. 2

Study Scope and Contributions

  • The benchmark evaluates executable pipelines across overall quality, error costs, thin-vessel retention, structural preservation, cross-dataset stability, and computational evidence, with equal dataset weighting.Results are intended for benchmark-relative selection, not architecture-only comparisons, because training exposure and implementation provenance differ.Source: §1 · p. 3
  • The ancillary analysis traces how vessel masks become centerline-, caliber-, and graph-based ring-wise features for downstream HRF classification.It also evaluates whether targeted feature stabilization reduces sensitivity to local mask errors.Source: §1 · p. 3
  • The study covers six evaluation datasets and five executable pipelines, totaling 244 images, and restricts segmentation evaluation to binary overall-vessel masks.Downstream classification is examined only as an ancillary HRF analysis.Source: §1 · p. 3

2 Evaluation Datasets and Included Pipelines

The benchmark uses heterogeneous public datasets and selects pipelines that can produce binary vessel masks through consistent executable evaluation pathways. Its dataset and pipeline choices are shaped by training exposure, provenance, and reproducibility.

2.1 Evaluation Datasets and Source Traceability

  • The benchmark includes 244 image–mask pairs from six public datasets spanning varied resolutions, acquisition conditions, pathology, and annotation procedures.The datasets provide heterogeneous domains for cross-dataset evaluation.Source: §2.1 · p. 4, Table 2.1 · p. 5
  • FIVES and DRIVE were excluded because evaluated pipelines had training exposure to those datasets.FIVES contributed training data for several pipelines and development of DCP-1024; GAVE used a DRIVE-trained checkpoint.Source: §2.1 · p. 4

2.2 Included Pipelines and Source Traceability

  • Five pipelines met the executable binary-mask and batch-prediction criteria; three were retrained locally, while DCP-1024 and GAVE used released checkpoints.Prompt-dependent methods were excluded, and the selected pipelines are not a comprehensive survey of the field.Source: §2.2 · pp. 4–5, Table 2.2 · p. 5

3 Pipeline Realization and Evidence Qualification

The benchmark preserves source-compatible operating procedures while recording configuration sensitivity and available computational evidence. Leading recommendations remained stable under tested configuration changes, but resource comparisons are incomplete.

3.1 Original Design and Source-Paper Operating Procedures

  • The pipelines differed in architecture, training, preprocessing, and inference settings, with detailed source-paper and released-implementation settings reported in an appendix.Source: §3.1 · p. 6

3.2 Benchmark Threshold Policy and Operating Points

  • Benchmark thresholds followed source implementations when specified and used 0.50 for SA-UNetv2 and DCP-1024 when source thresholds were unavailable.FSG-Net retained 0.50, FR-UNet retained dual-threshold iteration, and GAVE used a primary threshold of 0.35 with skeleton predictions at 0.10, 0.20, and 0.25.Source: §3.2 · p. 6
  • Threshold robustness checks found limited variation for FSG-Net, SA-UNetv2, and DCP-1024, but extreme sensitivity for GAVE; main selection conclusions were unaffected.Source: §3.2 · p. 6

3.3 Configuration Sensitivity and Selection Stability

  • Across tested hyperparameter settings, the principal recommendation pattern remained stable; FR-UNet varied more but did not surpass the leading pipeline.FSG-Net and SA-UNetv2 showed only limited cross-dataset changes within the examined ranges.Source: §3.3 · p. 6

3.4 Local Training and Computational Evidence

  • Among locally trained pipelines, SA-UNetv2 had the smallest model scale and lowest profiled forward-pass cost, while FR-UNet had the lowest recorded peak training VRAM.Resource comparisons remain provisional because DCP-1024 and GAVE lacked local training runs and several fields were unavailable.Source: §3.4 · p. 7, Table 3.1 · p. 7

4 Evaluation Metrics and Interpretation

Pixel agreement alone can miss structural error, so the benchmark interprets segmentation through complementary pixel-level, thin-vessel, and structural evidence. Its evaluation also considers cross-dataset robustness and aggregates dataset results with equal weighting.

4 Evaluation Metrics and Interpretation

  • Table 4.1 summarizes complementary metrics that assess segmentation quality from different perspectives.The benchmark distinguishes pixel-level, thin-vessel, and structural measures.Source: §4 · p. 7, Table 4.1 · p. 8

4.1 Limitations of Pixel-Based Evaluation

  • High Accuracy can arise from correct background classification even when no vessels are detected, so Accuracy is contextual rather than a primary model-selection criterion.Pixel-level F1 and IoU also do not identify the structural consequences of errors.Source: §4.1 · pp. 8–9
  • 07 dr and 15 g had nearly matched F1 (0.741571 and 0.740927), yet structural fidelity was 0.794865 versus 0.736642, with broader thin and peripheral false negatives in 15 g.Figure 4.2 distinguishes reference vessels in white, false negatives in black, and false positives in blue.Source: Figure 4.2 · p. 9, §4.1 · pp. 9–10

4.2 Required Multi-Dimensional, Task-Conditioned Evidence

  • The framework organizes evidence into pixel-level performance, cross-dataset robustness, thin-vessel retention, and structural fidelity for task-conditioned model selection.Pixel-level, thin-vessel, and structural measures are evaluated alongside cross-dataset stability.Source: §4.2 · p. 10

4.3 Evaluation Implementation and Aggregation Principles

  • Image-level metrics were averaged within each model-dataset pair, then datasets were equally weighted; cross-dataset standard deviation and worst-dataset performance described robustness.Structural fidelity used clDice and tolerant Skeleton F1; Thin-Vessel Sensitivity assessed recovery of ground-truth-defined thin branches.Source: §4.3 · p. 10

5 Evidence-Guided Model Selection

The framework combines benchmark evidence into task-conditioned pipeline recommendations rather than a universal ranking. Across the representative settings, FSG-Net and SA-UNetv2 led different objectives, while residual-error profiles add a diagnostic layer after shortlisting.

5.1.1 Evidence Utilities and Score Construction

  • The decision score combines pixel, thin-vessel, structural, and available resource utilities, aggregating dataset-level evidence equally with an optional worst-dataset risk term.The structural utility weights clDice and tolerant Skeleton F1; the prespecified η = 0.80 gives clDice the primary role, not as a clinical estimate or universal standard.Source: §5.1.1 · pp. 11–12

5.1.2 Scenario Presets and Benchmark-Relative Recommendations

  • No pipeline ranked first across all representative objectives: FSG-Net led pixel-level and structural-fidelity settings, while SA-UNetv2 led the other four presets.The presets were manually specified as interpretable starting points, and users can adjust evidence weights and cross-dataset risk in the public tool.Source: §5.1.2 · pp. 12–13, Table 5.1 · p. 13

5.1.3 Public Decision Tool and Interpretation Boundary

  • The browser-based tool recomputes scenario-specific rankings from fixed benchmark evidence and reports dimension-level utilities and residual-risk profiles, not a universal pipeline ranking.It supports custom weights and displays post-ranking residual-risk interpretation.Source: §5.1.3 · p. 13, Figure 5.1 · p. 14

5.2 Pixel-Level Performance and Cross-Dataset Robustness

  • Small observed gaps between leading pipelines are descriptive comparisons, not evidence of statistically significant differences across heterogeneous datasets.Image-level variability largely overlapped among FSG-Net, DCP-1024, and SA-UNetv2.Source: §5.2.1 · p. 15, §5.2.2 · p. 17

5.2.1 Overall Pixel Agreement and FN–FP Trade-offs

  • FSG-Net achieved the highest equal-dataset Macro-F1 (0.77) and IoU (0.63), while SA-UNetv2 had the highest mean Sensitivity (0.85).DCP-1024 had the highest mean Precision (0.77); FSG-Net and DCP-1024 tied for highest mean MCC (0.76).Source: Table 5.2 · p. 15, §5.2.1 · p. 15

5.2.2 Relative Dataset Response and Robustness

  • FSG-Net combined the highest equal-dataset Macro-F1 (0.77) with joint-lowest cross-dataset SD (0.05) and joint-best worst-case Macro-F1 (0.70).SA-UNetv2 shared the latter two results but had a slightly lower equal-dataset Macro-F1 (0.76).Source: Table 5.3 · p. 17, §5.2.2 · p. 17

5.3 Structure-level Performance and Model Selection

  • FSG-Net achieved the highest mean clDice (0.83) and SF1 (0.78), while SA-UNetv2 achieved the highest mean TVS (0.63) and mean Sensitivity (0.85).SA-UNetv2 achieved the highest TVS on five of six datasets; on HRF, its TVS was 0.01 below FSG-Net (0.53).Source: §5.3.1 · p. 18, Table 5.4 · p. 18, §5.3.2 · p. 19

5.4 Model-Specific Failure Profiles

  • Missed-vessel and extra-vessel errors together made up 65.60%–75.40% of classified structural-error pixels, with DCP-1024 showing more missed-vessel errors and SA-UNetv2 more extra-vessel errors.The five-category taxonomy distinguishes boundary shifts, width bias, missed vessels, and extra vessels; residual profiles can inform application-specific follow-up after shortlisting.Source: §5.4.1 · p. 20, Table 5.6 · p. 20, §5.4.2 · p. 22

6 Dataset Effects and Domain Dependence

Performance and residual-error patterns varied with dataset identity, and dataset-level differences often exceeded differences among pipelines. The six-dataset benchmark describes these domain-conditioned patterns but does not establish generalizable transfer relationships.

6.1 Dataset Heterogeneity

  • AFIO had high visible-field vessel occupancy (0.15) but a very low thin-skeleton share (0.90%), while HRF had relatively large vessel diameters and STARE relatively small diameters.Because the benchmark includes only six datasets, these descriptors were not used to estimate generalizable correlations or transfer relationships.Source: §6.1 · p. 22, Table 6.1 · p. 22

6.2 Dataset-Conditioned Performance and Variability

  • Dataset identity produced a broadly shared performance shift across pipelines; most datasets shifted in the same direction, although HRF was not fully consistent and its cross-pipeline spread stayed below 0.1.The model-centered comparison expresses performance relative to each pipeline’s own cross-dataset mean.Source: §6.2 · p. 23, Figure 6.1 · p. 23

6.3 Dataset-Conditioned Structural Performance

  • Across datasets, TVS ranged from 0.45 to 0.80 and Sensitivity from 0.57 to 0.95, exceeding the corresponding model-level ranges of approximately 0.47–0.63 and 0.77–0.85.HRF had the lowest mean SF1 (0.56) and relatively low TVS; AFIO had the lowest mean clDice (0.73), second-lowest mean SF1 (0.63), and consistently low TVS, Sensitivity, and Macro-F1.Source: Table 6.2 · p. 23, §6.3 · p. 24, Table 6.3 · p. 24

6.4 Dataset-Conditioned Error Profiles

  • Within datasets, pipelines generally had similar residual-error compositions, while dominant error patterns varied more across datasets than across pipelines.AFIO had more missed-vessel errors, DualModal2019 more too-thick and extra-vessel errors, STARE more boundary shifts, and HRF and LES mixed profiles; these patterns are not a validated transfer rule.Source: §6.4 · p. 24, Figure 6.2 · p. 25

7 Feature-Level Evaluation for Downstream Classification

This HRF proof of concept evaluates how segmentation masks shape ring-wise vascular features and downstream classification. Pixel-level rankings did not predict baseline classification ordering, and selective stabilization made FSG-Net and SA-UNetv2 the leading predicted-mask inputs, without establishing a universal pipeline ranking.

7.1 Motivation and Scope

  • The analysis treats each segmentation mask as an intermediate representation, converting it separately into ring-wise vascular features before classification.The classifier receives features generated from the corresponding mask.Source: §7.1 · pp. 25–26, Figure 7.1 · p. 26

7.2 Experimental Setting

  • On 45 HRF images across healthy, glaucoma, and diabetic retinopathy classes, each mask generated 11 ring-wise vascular features for feature-only leave-one-out classification.The vascular features were computed in nine concentric rings; an image-level classifier was retained only as contextual reference.Source: §7.2 · p. 26

7.3 Baseline Classification and Pixel–Task Mismatch

  • Baseline pixel-level segmentation rankings were not preserved in downstream classification, so mask similarity alone was insufficient to predict task performance.The GAVE–GT reversal and classifier mechanism are examined in Appendix J.2.Source: §7.3 · p. 27, Table 7.1 · p. 27

7.4 Controlled Perturbation Experiment

  • Perturbations of similar pixel magnitude produced different downstream effects depending on the vascular feature representation.The comparison distinguishes baseline train → perturbed test without retraining from perturbed LOOCV after classifier refitting.Source: §7.4 · p. 27, Table 7.2 · p. 27

7.5 Feature Stabilization

  • Stabilization targeted features with potential numerical amplification; five feature definitions were processed, while VLD remained unchanged because its spline variant did not improve sensitivity.The final scheme used spline centerlines, Gaussian-smoothed EDT sampling, and graph pruning selectively.Source: §7.5 · p. 28, Table 7.3 · p. 28

7.6 Classification After Stabilization

  • After selective stabilization, FSG-Net and SA-UNetv2 were the leading predicted-mask inputs for downstream vascular-feature classification on HRF.The authors describe this as a proof of concept, not a universal ranking; they call for investigation on independent datasets.Source: §7.6 · pp. 28–29, Table 7.4 · p. 28

8 Discussion, Limitations, and Conclusions

The benchmark favors FSG-Net and SA-UNetv2 for different application priorities, not a universal winner. Its recommendations are bounded by pipeline and dataset differences, benchmark-specific definitions, and incomplete resource evidence.

8 Discussion, Limitations, and Conclusions

  • The benchmark compares complete executable pipelines, not architectures under identical training conditions, and covers only five pipelines across six datasets and 244 images.Dataset-conditioned patterns are descriptive rather than validated transfer rules; final model and threshold selection still requires representative target-domain validation.Source: §8 · p. 29
  • Resource-oriented recommendations are provisional because comparable local training measurements covered only three pipelines and several fields were unavailable for DCP-1024 and GAVE.Source: §8 · p. 29
  • Benchmark-specific thin-vessel, skeleton-tolerance, and structural-error definitions are not universal, and alternative choices may change values when leading pipelines differ only slightly.Reference masks also leave uncertainty around faint peripheral vessels and exact boundaries, so boundary errors may reflect predictions, annotations, or both [21].Source: §8 · pp. 29–30
  • At consensus level 5, 61.5% of FP pixels and 37.2% of FN pixels fell within 20% of local vessel diameter from an annotated boundary.Excluding 10% and 20% boundary bands increased Macro-F1 for every pipeline without changing the tested ranking.Source: §8 · p. 30, Figure 8.1 · p. 30
  • Feature stabilization targets numerical amplification of mask errors, not conversion of imperfect masks into reference annotations, and excessive processing may suppress genuine vascular structure.The HRF downstream analysis is a proof of concept based on 45 images; larger cohorts and independent-dataset validation are needed before its ranking can be considered generalizable.Source: §8 · pp. 30–31
  • GT-informed oracle fusion raised Macro-F1 from 0.76 to 0.83 and IoU from 0.61 to 0.73, while Laplacian oracle correction raised mean F1 by 0.08–0.13 and mean IoU by 0.12–0.17.These diagnostic upper bounds motivate future GT-free fusion and image-guided correction, but are not deployable algorithms.Source: §8 · p. 31
  • FSG-Net and SA-UNetv2 were the strongest general candidates, but the preferred pipeline depends on error costs, structural needs, computational constraints, and target domain.FSG-Net was favored for pixel agreement or structural fidelity; SA-UNetv2 for thin-vessel recovery, downside risk, or resource-aware objectives.Source: §8 · pp. 29, 32
  • The conclusion extends evaluation beyond reference-mask agreement to whether segmentation preserves information needed by downstream biomedical analyses.The authors propose investigating structure-aware learning jointly with feature-stable representations; soft-clDice did not improve F1 under the tested SA-UNetv2 setting.Source: §8 · pp. 31–32

B Threshold Sensitivity and Benchmark Operating Points

A diagnostic analysis tested whether the retained benchmark operating points were in unusually sensitive regions of a common threshold response. Sensitivity varied by pipeline, but the principal model-selection conclusions remained stable under the tested variation.

B Threshold Sensitivity and Benchmark Operating Points

C Configuration Sensitivity and Selection Stability

Configuration sweeps produced only modest changes for SA-UNetv2 and FSG-Net, while FR-UNet showed greater sensitivity. Under the tested structural-weight sweep, FSG-Net remained first and SA-UNetv2 second.

C Configuration Sensitivity and Selection Stability

D Metric Definitions and Cross-Dataset Aggregation Rules

E.1 Sensitivity to the Structural Weight η

  • FSG-Net remained first and SA-UNetv2 second as η varied from 0.00 to 1.00, although their score gap narrowed with greater clDice weight.The structural-fidelity preset held ρ = 0.30 fixed; η did not change the recommendation.Source: §E.1 · p. 43, Figure E.1 · p. 43

E.2 Sensitivity to the Cross-Dataset Risk Weight ρ

Across the tested cross-dataset risk-weight sweep, most objective recommendations stayed unchanged, while the pixel-performance recommendation changed with ρ. Excluding boundary-adjacent pixels raised Macro-F1 for every pipeline but did not alter their rankings.

E.2 Sensitivity to the Cross-Dataset Risk Weight ρ

  • The mixed, low-FN/thin-vessel, structural-fidelity, and resource-oriented objectives retained their top-ranked pipeline across the ρ sweep; pixel performance switched from FSG-Net to SA-UNetv2 between ρ = 0.58 and 0.59.The table reports the first- and second-ranked pipelines at 0.1 intervals; mean-only and conservative unknown-domain settings represent distinct pixel-performance risk regimes.Source: §E.2 · p. 43, Table E.1 · p. 44

FSG-Net

  • The low-FN/thin-vessel objective used β = 2 and P/T/S/R = 50/50/0/0, while the structural-fidelity objective used η = 0.80 and P/T/S/R = 0/0/100/0.For objectives containing a structural component, η was fixed at 0.80.Source: FSG-Net · p. 44

F Sensitivity of Model Rankings to Boundary Uncertainty

  • Widening the excluded GT-boundary band from 0% to 20% increased Macro-F1 for all five pipelines, while their rankings remained unchanged.The analysis changed only the evaluated pixel region; score increases are not performance gains because each setting evaluates a different region.Source: F Sensitivity of Model Rankings to… · p. 44

G Error-Aware Oracle Vote Fusion Details

GT-informed oracle analyses show potential gains from combining local model votes or correcting known errors, but both rely on ground-truth information. The supplied clDice-loss feasibility discussion distinguishes that training experiment from clDice’s role as a structural benchmark metric.

G Error-Aware Oracle Vote Fusion Details

H Laplacian Oracle Repair of Known FN/FP Pixels

  • With known FN and FP locations, Laplacian oracle repair increased mean F1 by 0.08–0.13 and mean IoU by 0.12–0.17 across datasets.FN recovery ranged from 37.4% to 55.6%, and FP removal from 44.4% to 52.3%; at h = 1 and τ = 0, the method restored known FN pixels when Dh > τ and removed known FP pixels when Dh ≤ τ.Source: H Laplacian Oracle Repair of Known… · p. 46, Table H.1 · p. 46

I clDice-Augmented Loss Feasibility Experiment

  • The HRF soft-clDice-loss feasibility test addresses whether that training-loss addition improves the local result, not whether clDice remains useful as a structural benchmark metric.The passage cautions that the HRF test’s negative result is not evidence against using clDice as the principal structural evaluation metric.Source: I clDice-Augmented Loss Feasibility… · p. 46, Table I.1 · p. 47

J.1 Detailed Classifier Implementation and Image-Level Reference

The downstream analysis compares explicit vascular features with an image-level reference and traces how segmentation differences affect feature extraction and classification. Controlled perturbations reveal feature-specific sensitivity, and selective stabilization makes FSG-Net and SA-UNetv2 the leading predicted-mask inputs in the HRF analysis.

J.1 Detailed Classifier Implementation and Image-Level Reference

  • The vascular-only feature route makes predictions traceable to specific features, optic-disc-diameter rings, and properties such as vessel area, caliber, curvature, or graph connectivity.Masks are converted into skeleton/centerline, graph, and diameter representations before ring-wise features are supplied to the classifier.Source: §J.1 · pp. 47–48
  • On 45 HRF images, RETFound achieved LOOCV accuracy of 38/45 = 84.44%, while the complete GT-based feature set achieved 43/45 = 95.56%.The latter is an extended GT-based reference using a broader feature set than the vascular-only analysis; the study then restricts analysis to 11 mask-linked vascular features to isolate segmentation effects.Source: §J.1 · pp. 47–48

J.2 Detailed Explanation of Pixel–Classification Mismatch

  • GAVE’s unprocessed vascular-only accuracy was 34/45 versus GT’s 33/45, but after feature-specific stabilization the results reversed to GT 35/45 versus GAVE 25/45.The authors interpret the initial difference as a small-sample downstream feature effect, not evidence of superior GAVE segmentations; changing feature definitions altered the classifier representation.Source: §J.2 · p. 48
  • GT and predicted masks induce method-specific classifier dependencies: GT relied more on VAF, VCmean, and Dratio, while the other methods showed differing caliber, graph, and centerline-geometry dependence.The feature ranking reflects coefficient magnitudes in each mask-generated coordinate system, not segmentation quality; pixel rankings need not transfer to classification without reproducible class differences in feature extraction.Source: §J.2 · pp. 48–49, Table J.1 · p. 55

J.3 Detailed Mask, Skeleton, Graph, and Diameter Representations

  • A binary mask yields a centerline and 8-neighbor connectivity graph, while the Euclidean Distance Transform estimates local radius and diameter as 2 × EDT(x).Small boundary changes can remove thin branches, alter graph connections, or change diameter estimates; interior errors mainly affect VAF, while branch, boundary, and centerline errors can affect topology, caliber, curvature, and directional variation.Source: §J.3 · pp. 49–50, Figure J.1 · p. 50

J.4 Full Controlled Perturbation Protocol and Sensitivity Calculation

  • Controlled one-pixel dilation, one-pixel erosion, and five-pixel skeleton-tip pruning affected features differently: dilation and erosion mainly changed area, diameter, and thin-vessel statistics, while pruning more readily affected centerline geometry and graph topology.The networks were not retrained; perturbations were applied to GT masks or skeletons before feature re-extraction, with protocols for perturbed-test distribution shift and perturbed self-LOOCV after classifier re-adaptation.Source: §J.4 · pp. 50, 52

J.5 Stabilization, Parameter Settings, and Processing Effects

  • Selective stabilization targeted five features rather than uniformly smoothing all 11, using spline-fitted centerlines and smoothing of the continuous diameter map without moving the binary-mask boundary.The spline used smooth sigma px = 1.25; VLD was excluded because its mean sensitivity increased from 0.3055 to 0.3220 (-5.4%).Source: §J.5 · pp. 52–53

J.6 Complete Robust-Feature Classification and Coefficient Rankings

  • After selective stabilization, FSG-Net and SA-UNetv2 were the leading predicted-mask inputs on HRF; robust-5 accuracies were FSG 30/45 and SAUNet 29/45.The robust-5 analysis replaced only five feature definitions and retained the other six; paired changes were reported descriptively without an inferential test or multiple-comparison-adjusted p-value.Source: §J.6 · pp. 54–55

Related papers

Loading 2610.35357v1…