An Automated Framework for Large-Scale Graph-Based Cerebrovascular Analysis
Daniele Falcetta 1 , Liane S. Canas 2 , Lorenzo Suppa 1,3 , Matteo Pentassuglia 1 , Jon Cleary 2 , Marc Modat 2 , Sébastien Ourselin 2 , Maria A. Zuluaga 1,2 · 2025
Read An Automated Framework for Large-Scale Graph-Based Cerebrovascular Analysis by Daniele Falcetta 1 , Liane S. Canas 2 , Lorenzo Suppa 1,3 , Matteo Pentassuglia 1 , Jon Cleary 2 , Marc Modat 2 , Sébastien Ourselin 2 , Maria A. Zuluaga 1,2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-03869.txt.
Introduction
Blood vessels in the brain form a branching network whose shape may carry signatures of ageing and disease. Daniele Falcetta and colleagues present CaravelMetrics in a 2025 preprint, an automated framework that turns brain vessel scans into graphs. Skeletonised vessel centrelines become the graph's edges, and atlas-based regions divide the network into arterial territories.
From each graph the framework computes fifteen features spanning vessel shape, topology, fractal structure, and geometry, either for the whole network or region by region. Applied to hundreds of three-dimensional angiography scans from the IXI dataset, covering adults from their twenties to their eighties, the tool produced reproducible vessel graphs. The features tracked age and sex differences, and linked greater education with more complex vasculature, matching earlier literature. The authors offer the system for population-level studies of vascular health and normative modelling of ageing.
Catalog record
| CREATOR | Daniele Falcetta 1 , Liane S. Canas 2 , Lorenzo Suppa 1,3 , Matteo Pentassuglia 1 , Jon Cleary 2 , Marc Modat 2 , Sébastien Ourselin 2 , Maria A. Zuluaga 1,2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-03869
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Automated Framework for Large-Scale Graph-Based Cerebrovascular Analysis Daniele Falcetta 1 , Liane S. Canas 2 , Lorenzo Suppa 1,3 , Matteo Pentassuglia 1 , Jon Cleary 2 , Marc Modat 2 , Sébastien Ourselin 2 , Maria A. Zuluaga 1,2 Abstract We present CaravelMetrics , a computational framework for automated cerebrovascular analysis that models vessel morphology through skeletonization-derived graph representations. The framework integrates atlas-based regional parcellation, centerline extraction, and graph construction to compute fifteen morphometric, topological, fractal, and geometric features. The features can be estimated globally from the complete vascular network or regionally within arterial territories, enabling multiscale characterization of cerebrovascular organization. Applied to 570 3D TOF-MRA scans from the IXI dataset (ages 20–86), CaravelMetrics yields reproducible vessel graphs capturing age- and sex-related variations and education-associated increases in vascular complexity, consistent with findings reported in the literature. The framework provides a scalable and fully automated approach for quantitative cerebrovascular feature extraction, supporting normative modeling and population-level studies of vascular health and aging. † † address: 1 EURECOM, Sophia Antipolis, France 2 School of Biomedical Engineering & Imaging Sciences, King’s College London, UK 3 Politecnico di Torino, Torino, Italy 1 Introduction Brain blood vessels change throughout the lifespan, with vessel density declining, curvature increasing, and branching patterns reorganizing, with implications for brain health [ 1 ] . Understanding cerebrovascular morphology is critical for detecting disease and distinguishing pathological from normal changes. To that end, population-level analyses are crucial as individual vessel measurements cannot be interpreted without demographic-specific reference ranges. These reference ranges allow discrimination between healthy aging and pathological changes across groups, and identification of clinical biomarkers that distinguish disease-related alterations from normal variations within the population. However, large-scale analyses have been limited by manual vessel tracing and a lack of automated biomarker extraction tools that can be used at scale. Current approaches for cerebrovascular analysis can be classified into three categories. Segmentation-based methods, enabled by recent advances in deep learning for accurate automated vessel segmentation, typically from time-of-flight magnetic resonance angiography (TOF-MRA), estimate arterial volume from the segmented masks at the whole brain and regional levels [ 2 ] . Atlas-based approaches provide population references through statistical arterial atlases [ 3 ] , offering normative cerebrovascular anatomy baselines. However, both approaches focus on simple volumetric and diameter measurements that may fail to completely characterize the vascular tree. Finally, graph-based computational tools use a segmentation mask as input to generate a graph representation of the vascular tree, from which intensity and morphometric features are extracted [ 4 , 5 , 6 , 7 , 8 ] . While commonly used, these tools are either semi-automatic [ 4 , 7 ] or have been specifically designed for a particular type of image [ 5 , 6 , 8 ] , limiting their usage at scale for any neurovascular imaging modality. Figure 1: Overview of the vascular feature extraction framework. Table 1: List of vascular features extracted in within the vessel analysis framework. Category Metric Definition/Formulation Application Morphometric Total vessel length L=\sum_{(u,v)\in E}d_{uv} Structural integrity Volume |M|\,\Delta x\,\Delta y\,\Delta z Vascular volume Topological Bifurcation count N_{bif}=|\{v\in V:\deg(v)=3\}| Branching complexity Bifurcation density N_{bif}/L Number of loops |E_{k}|-|V_{k}|+1 Connectivity patterns Loop lengths \sum_{(u,v)\in\text{loop}}d_{uv} for each loop Abnormal degree nodes |\{v\in V:\deg(v)>3\}| Anomalous branching Connected components |\{G_{k}\}| Network integrity Fractal Fractal dimension -\frac{\Delta\log(N(\varepsilon))}{\Delta\log(\varepsilon)}; N(\varepsilon) the boxes at scale \varepsilon Multi-scale organization Lacunarity \frac{\text{Var}(n)}{[\text{Mean}(n)]^{2}}+1 ; n the number of voxels per box Spatial heterogeneity Geometric Geodesic length L_{geo}(P)=\sum_{i=1}^{n-1}d_{v_{i}v_{i+1}} Segment Arc Length Instantaneous curvature \kappa(t)=\frac{\|\mathbf{x}^{\prime}(t)\times\mathbf{x}^{\prime\prime}(t)\|}{\|\mathbf{x}^{\prime}(t)\|^{3}} on interpolated \mathbf{x}(t) Local geometry Mean curvature \bar{\kappa}=\frac{1}{L_{geo}}\int_{0}^{1}\frac{\kappa(t)}{n(t)}\|\mathbf{x}^{\prime}(t)\|dt Average tortuosity Mean square curvature \frac{1}{L_{geo}}\int_{0}^{1}\frac{\kappa^{2}(t)}{[n(t)]^{2}}\|\mathbf{x}^{\prime}(t)\|dt Curvature variability Arc-over-chord ratio L_{geo}(P)/\|\mathbf{x}_{v_{1}}-\mathbf{x}_{v_{n}}\| Segment tortuosity In this work, we present CaravelMetrics , an automated framework for analyzing the cerebrovascular tree. Our framework extracts fifteen morphometric, topological, fractal, and geometric vascular features from a graph generated from vascular binary masks, thus being agnostic to the underlying image modality. Similar to [ 2 ] , we integrate an atlas of arterial regions, allowing us to analyze the cerebrovascular tree at both the whole-brain and regional levels. We showcase the potential of our open-source tool by analyzing the IXI Dataset 1 1 1 https://brain-development.org/ixi-dataset/ , a cohort of 570 subjects aged 20 to 86 years whose cerebrovascular trees were imaged using TOF-MRA. Our analysis reveals that our features effectively capture variations in the vessels across a population. 2 Methods CaravelMetrics relies on representing the cerebrovascular network as a mathematical graph structure that is extracted from a vascular binary mask. The graph and the vascular binary mask are then used to extract morphometric, topological, fractal, and geometric features that can be aggregated at the whole-brain or regional level through the usage of an arterial atlas (Figure 1 ). Graph Representation of Cerebrovascular Networks. Let us denote M:\Omega\subset\mathbb{Z}^{3}\to\{0,1\} , a binary vascular mask obtained from any 3D imaging modality. Given M , we define its skeleton S=\operatorname{Skel}(M)\subseteq\Omega as a one-voxel-thick subset that preserves the topological properties of M , such that S\subseteq M , \operatorname{conn}(S)=\operatorname{conn}(M) , and \operatorname{hom}(S)=\operatorname{hom}(M) , where \operatorname{conn}(\cdot) denotes the number of connected components and \operatorname{hom}(\cdot) describes the topology (loops and holes). From the skeleton points \mathbf{x}=(x,y,z)\in S\subset\mathbb{Z}^{3} , we construct an undirected weighted graph G=(V,E) , where each skeleton point corresponds to a node v\in V with spatial coordinates \mathbf{x}_{v}\in\mathbb{Z}^{3} . Edges E=\{(u,v)\} connect pairs of nodes within a 3\times 3\times 3 neighborhood (26-connectivity), with edge weights defined as the Euclidean distance d_{uv}=\|u-v\|_{2} . A vascular graph G may consist of k disconnected subgraphs, each corresponding to an individual vessel component, such that G=\bigsqcup_{k}G_{k} . Within a given G_{k} , a root node serves as a reference point for segment extraction. We define three root nodes to ensure representative coverage of the vascular structure: R_{1} corresponds to the end-point ( \deg(v)=1 ) with the largest vessel diameter, R_{2} to the end-point with the second-largest diameter, and R_{3} to the bifurcation node ( \deg(v)\geq 3 ) with the largest diameter. If no bifurcation exists, R_{3} defaults to R_{1} . Based on this, a vessel branch or segment denotes the shortest path between a selected root node s\in\{R_{1},R_{2},R_{3}\} and each reachable end-point e\in V_{\text{end}} within the same G_{k} : P_{s\to e}=\arg\min_{p\in\mathcal{P}(s,e)}\sum_{(u,v)\in p}d_{uv}, (1) where \mathcal{P}(s,e) denotes the set of paths between s and e in G_{k} . Feature Extraction. The framework extracts fifteen complementary vascular features grouped into four categories (Table 1 ), each capturing a distinct property of cerebrovascular organization. Morphometric features quantify network extent and capacity through total vessel length and volume. Topological features describe branching architecture and connectivity using bifurcation count, loop number, and component analysis. Fractal features characterize hierarchical self-similarity via box-counting and lacunarity [ 9 ] . Finally, geometric features quantify vessel tortuosity, reflecting arterial stiffening and age-related alterations, by capturing local vessel morphology and shape irregularity through cubic-spline parameterization of discrete centerline segments (Eq. 1 ) as continuous curves \mathbf{x}(t):[0,1]\!\to\!\mathbb{R}^{3} , with curvature values weighted by path multiplicity n(t) and averaged across root nodes s\in\{R_{1},R_{2},R_{3}\} to obtain descriptors for each G_{k} . Atlas-Based Regional Analysis. Figure 2: Scanner-induced systematic bias. Consistent IOP shift across all features demonstrates acquisition artifacts. The cerebrovascular tree exhibits heterogeneous patterns across brain territories, motivating a region-based analysis across arterial territories. We enable such analysis by aligning an arterial