X-ray Coronary Angiography

Updated 31 January 2026

X-ray Coronary Angiography is a precise imaging modality that uses iodinated contrast and real-time X-ray fluoroscopy to visualize coronary vessel structures.
Advanced preprocessing and segmentation methods enhance vessel clarity, enabling quantitative stenosis assessment and integration with 3D/4D reconstruction techniques.
Integration of deep learning, geometric modeling, and temporal analysis in XCA drives improved diagnostic accuracy and guides interventional strategies for coronary artery disease.

X-ray Coronary Angiography (XCA) is a foundational imaging modality for the diagnosis, risk stratification, and procedural planning of coronary artery disease (CAD). In XCA, iodinated contrast medium is injected into the coronary arteries, and high-resolution, real-time 2D projections are acquired via a C-arm fluoroscopic system. The resulting image sequences delineate vessel luminal anatomy and allow for the detection, localization, and quantitative assessment of stenotic lesions, which underpin clinical decision-making and interventional strategies. The field has undergone rapid transformation due to the integration of advanced computational imaging, algorithmic vessel extraction, deep learning–based detection and segmentation, and innovations in 3D/4D reconstruction from highly undersampled or motion-contaminated datasets.

1. Imaging Principles and Clinical Workflow

XCA is performed by introducing a selective arterial catheter into the coronary ostia, followed by bolus administration of a radiopaque contrast agent and acquisition of serial X-ray projections at 7–30 frames per second. The spatial and temporal resolution of XCA enables visualization of stenoses, thrombotic occlusions, anomalous origins, and flow-limiting plaques. Standard clinical workflow includes:

Vascular access and catheter placement in the coronary ostium.
Injection of contrast and recording of fluoroscopic cine-loops at defined C-arm angulations (e.g., LAO/RAO, cranial/caudal).
Real-time 2D assessment of vessel patency, branching morphology, and dynamic flow patterns.
Manual or semi-automated measurement of minimum lumen diameter and reference diameter for stenosis quantification.
For advanced analysis or procedural planning, construction of 3D vessel models or functional assessment via pressure wires or computational fluid dynamics.

Key limitations of traditional 2D XCA include vessel overlap, foreshortening artifacts, loss of depth, and operator-dependent variability in lesion severity estimation (Popov et al., 2022, Bransby et al., 2023).

2. Preprocessing and Enhancement Techniques

Robust vessel visualization in XCA is impeded by low contrast, noise, non-uniform illumination, motion artifacts, and the presence of catheters and guidewires. Preprocessing serves to enhance vessel conspicuity, suppress background clutter, and provide standardized inputs for downstream analysis.

Morphological Transforms: Multi-scale top- and bottom-hat transforms (I_en = I + I_th - I_bh) are widely adopted to enhance tubular structures and suppress diffuse background (Mulay et al., 2021, Fazlali et al., 2017, Popov et al., 2022).
CLAHE and Adaptive Histogram Equalization: Enhance local dynamic range while mitigating over-amplification at tile boundaries. Frequently used in multi-channel preprocessing pipelines for state-of-the-art segmentation models (Hassan et al., 31 Oct 2025).
Homomorphic Filtering: Frequency-domain suppression of non-uniform illumination and glare using Butterworth high-pass filters; applied as a standard step in stenosis and tree segmentation (Lin et al., 2023, Liu et al., 2023).
Vesselness Filters: Hessian-based Frangi, Meijering, and Sato filters leverage eigen-analysis of the image's second derivative structure tensor at multiple scales to boost elongated, low-brightness features corresponding to vessel lumina (Yousefzadeh et al., 24 Jan 2026, Popov et al., 2022).
Temporal Background Subtraction: Principal component analysis (PCA), robust PCA (RPCA), and tensor robust principal component analysis (TRPCA) decompose cine sequences into slowly varying background (low-rank) and vessel (sparse) components (Fu et al., 2022, Qin et al., 2022). Spatiotemporal total-variation regularization further penalizes non-smoothness in the extracted vessel layer.

These enhancement methods are often composited in modern deep-learning pipelines as input channels or as adaptive preprocessors conditional on image content.

3. Vessel Segmentation and Stenosis Detection Algorithms

Segmentation of the coronary tree and detection/localization of stenosis are core tasks that enable quantitative analytics, lesion targeting, and procedural guidance.

Classic and Superpixel Methods

Multi-scale SLIC superpixels combined with vesselness probability measures and majority voting yield rapid, topology-preserving segmentations, further refined with ridge-based polishing and orthogonal-line boundary detection (Fazlali et al., 2017).
Classical pipelines are strengthened by learned per-image parameter prediction (e.g., SVR-tuned Frangi filter parameters), which boosts generalizability and transfer across dataset domains (Yousefzadeh et al., 24 Jan 2026).

Deep Learning Approaches

U-Net and Variants: Encoder–decoder architectures with skip connections (e.g., DenseNet121 encoders, Self-ONN decoders) achieve state-of-the-art accuracy in both internal and external validation cohorts (Hassan et al., 31 Oct 2025, Yousefzadeh et al., 24 Jan 2026). High-resolution FPN architectures with merged coronary+catheter supervision further stabilize performance under domain shift.
Attention and Transformer Blocks: CBAM, Swin Transformer, and especially Mamba state-space blocks have demonstrated competitive advantage in capturing long-range dependencies and modeling blurred, narrow vessels (Rostami et al., 2024).
Semantic Segmentation for Multi-class Structures: Networks incorporating gating modules or explicit multi-class heads have been shown to accurately parse vessels, catheters, and balloons, facilitating both anatomical separation and phase matching required for 3D reconstruction (Koland et al., 15 Sep 2025).
Style Transfer Networks: AdaIN-style models trained on natural image corpora achieve robust segmentation in a zero-shot fashion, bypassing the need for annotated XCA datasets (Mulay et al., 2021).
Joint Vessel-Type Labeling: Segmentation heads assigning identity to branches (RCA, LAD, LCX) enable anatomical analytics, such as vessel-specific metrics and region-based reporting (Yousefzadeh et al., 24 Jan 2026).
Object Detection Frameworks: Anchor-based detectors (e.g., RetinaNet, YOLOv8) deliver efficient bounding-box and pixel-wise localization of stenotic segments using fused DICE+BCE loss and graph-based logical validation to ensure anatomical consistency (Keshavarz et al., 2023, Liu et al., 2023, Rodrigues et al., 2021).

Performance: Reported Dice overlap coefficients exceed 0.91 on internal datasets for FPN models, with external domain drop to ≈0.80 unless modest fine-tuning is performed (Yousefzadeh et al., 24 Jan 2026). Mamba-based U-Net architectures achieve F1 scores approaching 0.69 on stenosis segmentation challenges, outperforming prior semi-supervised and transformer-only baselines (Rostami et al., 2024).

Algorithmic and Practical Advances

Real-time inference rates (≈10–20 fps on modern GPUs) are achievable for most leading architectures.
Accurate, continuous vessel masks are critical for reliable quantification of percentage diameter stenosis, automated SYNTAX scoring, and integration with downstream 3D reconstruction workflows.

4. 3D and 4D Coronary Tree Reconstruction

Addressing the fundamental limitations of 2D projection, XCA-based vessel reconstruction methods leverage both geometric and deep-learning-based frameworks:

Geometric Approaches

Viewpoint Planning: Closed-form algorithms compute optimal secondary C-arm orientations to minimize vessel foreshortening by calculating rotations about user-selected coronary segments, ensuring at least 30° true rotation for robust quantitative coronary angiography (QCA) (Preuhs et al., 2018).
Biplanar Triangulation: Matching anatomical landmarks (branch/endpoints) between phase-matched XCA frames, geometric triangulation (with or without NURBS surface fitting) yields submillimeter-accuracy 3D centrelines (mean back-projection error ≈0.62 mm) (Koland et al., 15 Sep 2025).
Mesh-based Deep Learning: GCN-based architectures such as 3DAngioNet combine EfficientB3-UNet segmentations with graph deformation of mesh templates, outputting full 3D models with mean absolute error <0.35 mm versus clinical ground truth (Bransby et al., 2023).

Neural Implicit and NeRF-based Fields

Implicit Occupancy Networks: NeCA reconstructs 3D coronary trees from two 2D projections using multiresolution hash encoding and a differentiable cone-beam forward projector, trained with self-supervised loss on projection-consistency only. Dice overlap >90% (RCA) and mean surface error ≈0.14 mm have been achieved, with strong topology preservation (Wang et al., 2024).
Generative Models for Non-simultaneous Views: DeepCA combines a 3D U-Net generator (with latent transformers) and a dynamic snake-convolution critic to reconstruct topologically correct arterial trees even when the two projections are misaligned due to cardiac/respiratory motion, reporting Chamfer-L2 distances <4.51 mm on clinical ICA (Wang et al., 2024).
4D Dynamic Fields: NeRF-CA and NerT-CA extend implicit fields to time-varying settings, enabling dynamic, temporally resolved reconstructions from as few as three to four angiograms by modeling the foreground as a dynamic neural field atop a static background tensorial field. NerT-CA in particular achieves over 10× speedup compared to prior NeRF-based pipelines (train time ≈37 min), with Dice ≈0.75–0.88 for XCAT/MAGIX digital phantoms in 3-view scenarios (Maas et al., 25 Jul 2025, Maas et al., 2024).

Limitations and Open Problems

Reconstruction accuracy in real-world settings is constrained by unmodeled detector scatter, varying contrast kinetics, geometric miscalibration, and, in single-plane settings, the inherent ill-posedness of back-projecting from two views only. Most high-fidelity algorithms are currently validated primarily on phantom or simulated projections (Wang et al., 2024, Maas et al., 25 Jul 2025).

5. Temporal Analysis, Video Processing, and Dose Reduction

XCA cine sequences enable temporal modeling:

Frame Selection: Precision frame selection based on histogram/contrast criteria or deep sequence classifiers (CNN+LSTM, U-Net) directly impacts segmentation/detection fidelity by avoiding motion-blurred or partially contrasted frames (Popov et al., 2022, Yousefzadeh et al., 24 Jan 2026).
Robust PCA and Tensor Decomposition: Spatiotemporal decomposition (TRPCA, RPCA-UNet) segregates moving vessels from quasi-static background. Patch-wise super-resolution and CLSTM modules deliver SNR gains especially for distal branches, with global CNR improvements from ≈1.06 to 1.78 (Qin et al., 2022, Fu et al., 2022).
Deep Video Interpolation: Retrained DAIN models synthesize high-frame-rate cine sequences from lower-rate acquisitions, enabling 50% dose reduction with clinical PSNR ≈34 dB in interpolated frames. Such real-time frame synthesis preserves visual continuity while reducing cumulative X-ray exposure by up to 65% for both patient and operator (Yin et al., 2020).

6. Clinical Integration, Limitations, and Future Directions

Automation and analytics in XCA have direct clinical impact:

Segmentation and Lesion Mapping: Pixel- and region-level stenosis maps aid interventional planning, precise stent sizing, and longitudinal monitoring (Lin et al., 2023, Rostami et al., 2024, Rodrigues et al., 2021).
Vessel-Type Assignment: Per-pixel classification supports vessel-specific measurement and computational flow modeling, achieving >98% accuracy for RCA and ≈0.79 Dice for LAD/LCX (Yousefzadeh et al., 24 Jan 2026).
Procedural Guidance: Real-time segmentation and overlay onto live fluoroscopy may assist guidewire navigation and lesion crossing (Mulay et al., 2021, Hassan et al., 31 Oct 2025).
3D/4D Furniture: Automated tree reconstruction supports non-invasive FFR computation, planning of revascularization, and education.

Challenges: Major open issues remain, including sensitivity to image quality variation, generalizability under domain shift, incomplete labeling of minor branches, the need for large (preferably multi-center, multi-vendor) datasets, and extremely low SNR in challenging clinical scenarios (Popov et al., 2022, Hassan et al., 31 Oct 2025).

Future avenues: Topology-aware or skeletonization losses for 3D models, explicit modeling of cardiac and respiratory motion, polyenergetic/projector physics in rendering layers, and rapid, self-supervised adaptation across centers. Integration with clinical point-of-care systems hinges on regulatory-grade explainability, tight latency constraints (<100 ms/frame), and modular pipelines for staged verification and quality control.

Key References

(Mulay et al., 2021) Style Transfer based Coronary Artery Segmentation in X-ray Angiogram
(Lin et al., 2023) StenUNet: Automatic Stenosis Detection from X-ray Coronary Angiography
(Liu et al., 2023) YOLO-Angio: An Algorithm for Coronary Anatomy Segmentation
(Rostami et al., 2024) Segmentation of Coronary Artery Stenosis in X-ray Angiography using Mamba Models
(Koland et al., 15 Sep 2025) 3D Reconstruction of Coronary Vessel Trees from Biplanar X-Ray Images Using a Geometric Approach
(Hassan et al., 31 Oct 2025) CASR-Net: An Image Processing-focused Deep Learning-based Coronary Artery Segmentation and Refinement Network for X-ray Coronary Angiogram
(Maas et al., 25 Jul 2025) NerT-CA: Efficient Dynamic Reconstruction from Sparse-view X-ray Coronary Angiography
(Maas et al., 2024) NeRF-CA: Dynamic Reconstruction of X-ray Coronary Angiography with Extremely Sparse-views
(Yousefzadeh et al., 24 Jan 2026) Coronary Artery Segmentation and Vessel-Type Classification in X-Ray Angiography
(Fu et al., 2022) Robust Implementation of Foreground Extraction and Vessel Segmentation for X-ray Coronary Angiography Image Sequence
(Qin et al., 2022) Robust PCA Unrolling Network for Super-resolution Vessel Extraction in X-ray Coronary Angiography
(Bransby et al., 2023) 3D Coronary Vessel Reconstruction from Bi-Plane Angiography using Graph Convolutional Networks
(Wang et al., 2024) NeCA: 3D Coronary Artery Tree Reconstruction from Two 2D Projections via Neural Implicit Representation
(Popov et al., 2022) A Review of Modern Approaches for Coronary Angiography Imaging Analysis
(Yin et al., 2020) Reducing the X-ray radiation exposure frequency in cardio-angiography via deep-learning based video interpolation
(Preuhs et al., 2018) Viewpoint Planning for Quantitative Coronary Angiography