ArteryX: Semi-Supervised Artery Analysis

• ArteryX is a MATLAB-based, semi-supervised framework that quantifies cerebral artery structure, geometry, morphology, and topology from 3D TOF MRA.
• It integrates a vessel-fused network with landmark-guided tracing to efficiently manage both proximal and distal arterial segments.
• The framework employs an in-vivo simulation for quantitative validation, reducing user intervention and enhancing sensitivity to subtle vascular changes.

Searching arXiv for the specified ArteryX paper and closely related work for contextual support. ArteryX is a MATLAB-based semi-supervised framework for artery-specific feature extraction from three-dimensional Time-of-Flight Magnetic Resonance Angiography (3D TOF MRA), developed to quantify intracranial arterial structure, geometry, morphology, and topology with limited user intervention while introducing a stronger validation paradigm than the weakly labeled evaluations common in prior work (Faiyaz et al., 10 Jul 2025). It is designed for cerebral vasculature analysis in settings where subtle vascular alterations, especially beyond the major proximal vessels of the Circle of Willis, may be clinically informative but are not routinely quantified in standard radiological workflows (Faiyaz et al., 10 Jul 2025). A distinctive aspect of the framework is the combination of a vessel-fused network based landmarking approach for robust tracing management and an in-vivo like simulation framework with predefined artery-specific ground-truth features for quantitative validation (Faiyaz et al., 10 Jul 2025).

1. Definition, scope, and clinical rationale

ArteryX targets the extraction of artery-specific and network-level cerebrovascular features from 3D TOF MRA, including metrics such as artery radius or diameter, length, tortuosity, branch count, surface area, volume, and complexity or topology (Faiyaz et al., 10 Jul 2025). The framework is motivated by the observation that routine clinical evaluation of TOF MRA generally emphasizes major abnormalities such as stenosis or aneurysm, whereas more distributed and subtle vascular changes may remain unmeasured despite their potential relevance to cerebrovascular pathology, cognitive decline, and neurological disorders (Faiyaz et al., 10 Jul 2025).

The stated application domain includes proximal arteries of the Circle of Willis and distal arterial subnetworks, particularly the MCA, ACA, and PCA territories (Faiyaz et al., 10 Jul 2025). This emphasis on distal vasculature is important because the paper argues that distal artery quantification is often ignored due to methodological difficulty and variability, even though such features may be sensitive to early vascular remodeling or rarefaction (Faiyaz et al., 10 Jul 2025). The framework is presented in the context of cerebral small vessel disease, hypertension, diabetes, dyslipidemia, smoking, HIV, chronic kidney disease, obesity or metabolic syndrome, obstructive sleep apnea, autoimmune disease, and radiation exposure or chemotherapy (Faiyaz et al., 10 Jul 2025).

The paper positions ArteryX against three limitations in prior approaches. First, manual or weakly supervised assessments exhibit inter- and intra-rater variability (Faiyaz et al., 10 Jul 2025). Second, fully automated and deep learning methods often require large training datasets and may remain protocol-specific or focused mainly on segmentation or proximal structures (Faiyaz et al., 10 Jul 2025). Third, existing semi-automated methods can be labor-intensive, especially when dangling or disconnected vessels require substantial cleanup (Faiyaz et al., 10 Jul 2025). This motivates a semi-supervised design in which limited user landmarking is combined with automated graph-based processing (Faiyaz et al., 10 Jul 2025).

2. Input data and acquisition assumptions

ArteryX is designed for 3D TOF MRA of the intracranial vasculature (Faiyaz et al., 10 Jul 2025). For the in-vivo cohort comparison reported in the paper, imaging was performed on a Siemens 3T MAGNETOM Prisma Fit with a 64-channel head coil and gradients of 80 mT/m maximum strength and 200 mT/m/s slew rate (Faiyaz et al., 10 Jul 2025). The TOF MRA sequence parameters reported were $TE/TR = 3.42/21 \, \text{ms}$, flip angle $= 18^\circ$, and spatial resolutions of $0.52 \times 0.52 \times 0.8 \, \text{mm}^3$ and $0.26 \times 0.26 \times 0.5 \, \text{mm}^3$ (Faiyaz et al., 10 Jul 2025).

The abstract states that ArteryX achieves processing times of approximately 10–15 minutes per subject at 0.5 mm resolution with minimal user intervention (Faiyaz et al., 10 Jul 2025). The paper further states that user interaction in routine use is primarily limited to landmarking up to 16 main nodes in a 3D viewer, typically requiring about 8–10 minutes per subject (Faiyaz et al., 10 Jul 2025). This is contrasted with the comparator iCafe, which required landmarking plus correction of dangling or disconnected arteries and took about 1.5 hours per subject (Faiyaz et al., 10 Jul 2025).

The framework assumes that the vascular image can be transformed into an arterial binary mask and then into an isotropic representation for feature measurement (Faiyaz et al., 10 Jul 2025). Images were also checked for motion artifacts and signal dropout (Faiyaz et al., 10 Jul 2025).

3. Computational pipeline

ArteryX is described as a multi-stage framework whose core operations are artery segmentation, centerline extraction, isotropic rasterization, radius estimation, graph construction, landmark-guided artery classification, tracing management, and feature quantification (Faiyaz et al., 10 Jul 2025). The paper summarizes the workflow in Algorithm 1 (Faiyaz et al., 10 Jul 2025).

The first stage is unsupervised artery segmentation using Expectation-Maximization within a Hidden Markov Random Field framework, denoted HMRF-EM (Faiyaz et al., 10 Jul 2025). This stage separates artery voxels from surrounding tissue using voxel intensities, Gaussian class models, spatial regularization via a Markov random field, and iterative EM updates with ICM (Faiyaz et al., 10 Jul 2025). This is a fully unsupervised segmentation step rather than a learned supervised vessel mask predictor (Faiyaz et al., 10 Jul 2025).

The second stage is centerline extraction from the artery binary image (Faiyaz et al., 10 Jul 2025). The framework computes a one-voxel-thick 3D arterial skeleton using topology-preserving thinning, with the intention of retaining branches and connectivity (Faiyaz et al., 10 Jul 2025). This skeleton forms the backbone for later tracing and graph-based artery management (Faiyaz et al., 10 Jul 2025).

The third stage addresses voxel anisotropy and local caliber measurement (Faiyaz et al., 10 Jul 2025). The artery image is resampled into an isotropic grid, and a Euclidean distance transform of the inverted binary vessel image is used to estimate the distance from each centerline voxel to the nearest vessel wall, which becomes the local radius (Faiyaz et al., 10 Jul 2025). This suggests a centerline-attached radius field over the vascular graph.

The framework then constructs a vessel-fused network representation, using graph nodes and tracing logic to manage connectivity and classification (Faiyaz et al., 10 Jul 2025). The paper states that vessel-fused network based landmarking reliably tracks and manages tracings and effectively addresses the issue of dangling or disconnected vessels (Faiyaz et al., 10 Jul 2025). The key user interaction occurs here: the user labels up to 16 critical landmarks sequentially in a 3D viewer with Hessian guidance, and spurious edges can be deleted if needed in a labeling panel (Faiyaz et al., 10 Jul 2025). The framework standardizes classification with 16 critical nodes spanning artery types such as ICA, MCA, ACA, PComm, PCA, BA, and VA, including left-right specific nodes where appropriate (Faiyaz et al., 10 Jul 2025).

After tracing and classification, ArteryX quantifies artery-specific and network-level features (Faiyaz et al., 10 Jul 2025). These include radius or diameter, length, tortuosity, branch count, surface area, volume, and complexity or topology measures (Faiyaz et al., 10 Jul 2025). The paper’s emphasis is not only on proximal labeled segments but also on distal subnetworks, especially MCA, ACA, and PCA trees (Faiyaz et al., 10 Jul 2025).

4. Vessel-fused network based landmarking and tracing management

The vessel-fused network based landmarking strategy is the method’s central algorithmic contribution (Faiyaz et al., 10 Jul 2025). The paper argues that tracing cleanup is a major bottleneck in previous semi-automated tools because dangling or disconnected vessels require extensive manual intervention (Faiyaz et al., 10 Jul 2025). ArteryX addresses this by integrating landmarking directly with a vessel-fused graph structure, allowing tracings to be managed in a way that is robust to disconnected or dangling segments (Faiyaz et al., 10 Jul 2025).

User supervision is deliberately narrow in scope. The operator identifies up to 16 main landmarks in a 3D viewer, sequentially assisted by a Hessian guide, and can remove spurious edges if necessary (Faiyaz et al., 10 Jul 2025). Once these landmarks are placed, the graph representation organizes artery tracing, classification, and feature attribution (Faiyaz et al., 10 Jul 2025). The paper presents this as a substantial usability improvement over iCafe, in which dangling or disconnected vessels remain a major source of manual effort (Faiyaz et al., 10 Jul 2025).

A plausible implication is that the “vessel-fused” representation serves not only as a classification scaffold but also as a topology stabilizer for downstream measurements. The paper explicitly states that the approach improves handling of dangling or disconnected vessels (Faiyaz et al., 10 Jul 2025), and this directly affects the reliability of branch-based and distal-vessel features.

5. Quantified arterial features

ArteryX is designed to quantify a range of structural, geometrical, morphological, and topological features from intracranial arteries (Faiyaz et al., 10 Jul 2025). The paper explicitly mentions artery radius or diameter, length, tortuosity, branch count, surface area, volume, and complexity or topology (Faiyaz et al., 10 Jul 2025). These measurements are intended for artery-specific analysis as well as network-level characterization (Faiyaz et al., 10 Jul 2025).

The framework is presented as especially useful for subtle vascular change detection rather than only major lesion detection (Faiyaz et al., 10 Jul 2025). The paper’s motivation is that diffuse or distal vascular abnormalities may manifest through morphology and topology, not only through obvious focal stenosis or aneurysm (Faiyaz et al., 10 Jul 2025). This makes the framework relevant for cohort comparison studies and diseases in which distributed vascular alterations are hypothesized (Faiyaz et al., 10 Jul 2025).

This feature-centric emphasis distinguishes ArteryX from work that focuses only on segmentation or vessel-type classification. For example, recent angiography systems prioritize segmentation and artery labeling in X-ray coronary angiography (Yousefzadeh et al., 24 Jan 2026), whereas ArteryX’s stated novelty lies in artery-specific feature extraction and quantitative validation on MRA (Faiyaz et al., 10 Jul 2025).

6. Validation framework

A major claim of the paper is that ArteryX introduces a more rigorous validation strategy than the weakly labeled assessments common in earlier vessel-analysis tools (Faiyaz et al., 10 Jul 2025). The framework includes an in-vivo like artery simulation environment that uses vessel-fused graph nodes and predefined ground-truth features for specific artery types (Faiyaz et al., 10 Jul 2025). This simulation capability is presented as enabling quantitative feature validation rather than relying solely on manual annotations, which are themselves variable (Faiyaz et al., 10 Jul 2025).

The paper explicitly states that this simulation framework is designed to benchmark feature extraction toolboxes and standardize comparisons across patient cohorts (Faiyaz et al., 10 Jul 2025). This is one of the paper’s defining methodological points: ArteryX is not only a measurement tool but also a validation framework for artery feature extraction (Faiyaz et al., 10 Jul 2025).

This validation orientation is unusual in the artery-analysis literature summarized in the broader dataset. For instance, lower-extremity CTA approaches such as the rule-based system of Wilson et al. emphasize slice-wise vessel counting and calcification or stenosis profiles but lack strong expert-labeled lesion validation (Zhao et al., 2019). Likewise, many segmentation systems report Dice or overlap metrics but do not provide explicit artery-type feature ground truth through simulation (Rajeoni et al., 2023, Zeng et al., 31 Jul 2025, Yousefzadeh et al., 24 Jan 2026). ArteryX’s simulated ground-truth feature framework therefore occupies a distinct methodological niche (Faiyaz et al., 10 Jul 2025).

7. Comparative performance and workflow efficiency

The paper reports that ArteryX demonstrated improved sensitivity to subtle vascular changes and better performance than an existing semi-automated method in human subjects with cerebral small vessel disease (Faiyaz et al., 10 Jul 2025). The comparator emphasized in the text is iCafe (Faiyaz et al., 10 Jul 2025). The paper also reports a major workflow advantage: approximately 10–15 minutes per subject at 0.5 mm resolution, with only 8–10 minutes of user interaction, compared with approximately 1.5 hours per subject for iCafe (Faiyaz et al., 10 Jul 2025).

These gains are framed not primarily as segmentation gains but as improvements in feature extraction efficiency, tracing management, and standardization (Faiyaz et al., 10 Jul 2025). The paper’s evaluation focus is aligned with downstream quantitative vascular phenotyping rather than only voxelwise overlap.

A plausible implication is that ArteryX is especially suited to medium-scale cohort studies where semi-automated accuracy is desired but full manual tracing is prohibitively slow. The paper explicitly links the framework to seamless integration into clinical workflows and standardized comparisons across patient cohorts (Faiyaz et al., 10 Jul 2025).

8. Relation to adjacent artery-analysis research

ArteryX occupies a different methodological space from several contemporaneous artery-analysis systems. In retinal imaging, loss-centric deep learning systems couple vessel, artery, and vein outputs to improve anatomical consistency and fine-grained classification (Zeng et al., 31 Jul 2025). In X-ray coronary angiography, segmentation pipelines emphasize catheter-aware supervision, vessel-type labeling, and external transfer across centers (Yousefzadeh et al., 24 Jan 2026), or image-enhancement and topology-preserving decoders for stenotic branch continuity (Hassan et al., 31 Oct 2025). In CTA of peripheral arteries, deep learning systems combine artery segmentation with calcification scoring in a vessel-first workflow (Rajeoni et al., 2023). ArteryX instead centers on 3D TOF MRA, vessel-fused landmarking, distal intracranial feature extraction, and simulation-based quantitative validation (Faiyaz et al., 10 Jul 2025).

This suggests that ArteryX is best understood not as a generic artery segmentation network, but as a cerebrovascular phenotyping framework whose distinctive contributions are graph-guided artery management and validation methodology (Faiyaz et al., 10 Jul 2025). The broader literature indicates that many artery-analysis pipelines stop at segmentation or simple biomarker derivation, whereas ArteryX formalizes artery-specific feature extraction and benchmarking in a single toolbox (Faiyaz et al., 10 Jul 2025).

9. Limitations and interpretation

ArteryX is explicitly semi-supervised rather than fully automated (Faiyaz et al., 10 Jul 2025). User interaction remains necessary for landmarking and occasional edge deletion (Faiyaz et al., 10 Jul 2025). The framework is also MATLAB-based (Faiyaz et al., 10 Jul 2025), which may be relevant for integration and deployment considerations, though the paper presents this as an implementation choice rather than a methodological limitation.

The framework is designed around 3D TOF MRA of intracranial arteries (Faiyaz et al., 10 Jul 2025). The data block does not state that it has been generalized to other vascular territories or imaging modalities. Any such generalization would therefore be an inference rather than an established result. Similarly, while the paper claims improved sensitivity to subtle vascular changes in human subjects with cerebral small vessel disease (Faiyaz et al., 10 Jul 2025), the exact cohort size and full statistical breakdown are not present in the provided block and therefore cannot be specified further here.

The simulation-based validation framework is a major strength, but the paper’s abstract and details do not fully specify all simulated artery classes, noise models, or generative assumptions in the available excerpt (Faiyaz et al., 10 Jul 2025). This suggests that the framework’s validation philosophy is clear, while complete implementation detail likely resides in the full paper and toolbox documentation.

10. Significance

ArteryX is significant as a cerebrovascular analysis framework because it shifts emphasis from visual inspection of major lesions toward quantitative intracranial artery phenotyping and because it couples semi-automated graph-based artery management with a robust validation strategy (Faiyaz et al., 10 Jul 2025). Its vessel-fused network based landmarking addresses a practical failure mode of prior semi-automated tools—dangling and disconnected vessels—while its in-vivo like simulation framework addresses a methodological weakness in prior validation practices (Faiyaz et al., 10 Jul 2025).

The framework therefore sits at the intersection of artery segmentation, graph-based vascular modeling, distal vessel quantification, and benchmarking methodology. Its reported 10–15 minute per-subject processing time at 0.5 mm resolution and limited user interaction suggest that it is intended not only for research prototyping but also for eventual workflow integration in cohort studies and clinically oriented vascular assessment (Faiyaz et al., 10 Jul 2025).