An arXiv preprint reports that one 3D vessel-segmentation model led the paper’s reported comparisons across three public medical-imaging datasets. 3D-CurvSegFlow had the best Dice, centerline Dice (clDice) and recall on every dataset. It also had the best intersection over union (IoU), precision and 95th-percentile Hausdorff distance (HD95) on two of the three. That is a descriptive benchmark result: it shows how the model ranked on the paper’s measures, not whether it improves diagnosis, treatment planning or patient outcomes.
The work is an arXiv preprint, version 1, dated 20 Aug 2026. Its central question is whether a unified conditional flow-matching model can achieve robust performance across diverse 3D curvilinear structures. The same architecture and identical training strategy were used across all datasets, without anatomy-specific modifications. That makes the report a test of one shared setup rather than a set of individually adapted systems.
One architecture across three datasets
3D-CurvSegFlow uses a 3D time-conditioned U-Net with attention-gated skip connections. In practical terms, its flow-matching setup moves toward a segmentation through a short sequence of integration steps: at inference—the stage when the trained network produces its output—the reported system uses N = 3 steps followed by a sigmoid. Training combines the terms denoted LFM, LWBCE and LDice, corresponding to flow-matching, weighted binary cross-entropy and Dice loss.
3Dircadb contributed 19 contrast-enhanced abdominal CT scans, with a patient-wise split of 15 training volumes and 4 test volumes. SMILE-UHURA began as a collection of 20 original subjects, but the reported experiment had 14 available labeled samples; 12 were used for training and 2 for testing. The difference between the original and available counts is important context when the results are compared across datasets.
ImageCAS comprised 1000 CTA volumes from unique patients, split into 800 training and 200 testing volumes. The reported test sets therefore differed sharply in size, from 4 volumes in 3Dircadb and 2 in SMILE-UHURA to 200 in ImageCAS. Yet the model’s architecture and training strategy were kept common across the datasets.
What the tables show
In the reported comparisons, 3D-CurvSegFlow had the best Dice, clDice and recall across all three datasets. It had the best IoU, precision and HD95 on two of them. The pattern spans several different measures rather than a single headline score, but the comparison remains a descriptive ranking within the reported datasets and comparator settings.
On 3Dircadb, the proposed model’s row reported Dice 0.74, IoU 0.587, precision 0.78, recall 0.71 and clDice 0.70. Its final listed metric was 18.54. These are the reported outputs for the portal-vein task, not a measure of patient benefit.
On SMILE-UHURA, the proposed model’s row reported Dice 0.8032, IoU 0.6763, precision 0.8780, recall 0.7734 and clDice 0.8418. The table’s final listed metric was 3.87. The result is one part of the cross-dataset comparison, alongside the separate 3Dircadb and ImageCAS evaluations.
On ImageCAS, the proposed model’s row reported Dice 0.824, IoU 0.701, precision 0.8020, recall 0.850 and clDice 0.8793. Its final listed metric was 22.62. The reported ranking therefore rests on performance across the three tables rather than on the ImageCAS result alone.
Speed becomes part of the story
The paper also reports inference time—the time required to process one volume once the model is producing a segmentation. On SMILE-UHURA, the reported inference result was 22.3 seconds per volume with a general Dice score of 0.8032. The 5-step run took 36.4 seconds and had Dice 0.773, while the 10-step run took 71.9 seconds and had Dice 0.753.
Those step-count figures do not establish that 3 steps are globally optimal. They describe the reported SMILE-UHURA comparison: the 5-step and 10-step runs were slower and had lower reported Dice values than the 3-step result. The supplied analysis identifies alternative ODE solvers and adaptive integration as open ways to examine the balance between speed and accuracy.
Where the evidence stops
The authors interpret the cross-dataset rankings and qualitative examples as suggesting robust generalization, preservation of vascular continuity and topology, and computational efficiency. The study design supports descriptive benchmark comparisons within the tested public datasets, not clinical-effectiveness conclusions. It does not show improved diagnosis, treatment planning, surgery or patient outcomes.
The imaging coverage also leaves important questions open. Only contrast-enhanced CT and high-resolution TOF-MRA datasets were evaluated; low- or non-contrast modalities remain untested. Fixed-size volumetric patches may restrict global anatomical context for very large vascular trees. The analysis also notes occasional distal over-segmentation on coronary CTA affecting HD95.
Comparisons with baselines need caution because comparator results include values from prior studies, a few-shot task and 5-fold cross-validation, so the settings are not fully equivalent. The supplied analysis reports no confidence intervals, p-values or formal significance tests. The tables show the reported rankings, but not how much statistical uncertainty surrounds their differences.
Taken together, the strongest conclusion is a technical one: a single common model ranked highly on the paper’s chosen measures across several vessel datasets. The report does not establish transfer to larger datasets, additional vascular territories, low- or non-contrast imaging, or unseen clinical settings. It also leaves open whether alternative integration methods, broader context and better coronary boundary handling would change the speed–accuracy balance.
Paper data and sources
Original title: Flow Matching Meets 3D Curvilinear Structure Segmentation in Medical Imaging
Authors: Sidi Mohamed Sid'El Moctar, Nicolas Vitry, Hélène Bouvrais
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text