AngioAI-QFR: Automated Coronary Physiology Analysis

• AngioAI-QFR is an end-to-end angiography-only pipeline that integrates deep learning for lesion detection and segmentation with 1D hemodynamic modeling to compute QFR.
• The system automates lumen segmentation, RFC profiling, and virtual stenting simulation, achieving 93% automatic execution and showing strong agreement with invasive FFR (r=0.89).
• Its robust performance, sub-minute runtime, and accurate physiology assessment support its practical use in guiding PCI planning in coronary artery disease.

AngioAI-QFR is an end-to-end, angiography-only pipeline designed to automate the assessment of coronary artery disease, combining deep learning–based lesion detection and segmentation, centerline and diameter extraction, per-millimeter Relative Flow Capacity (RFC) profiling, virtual stenting with automatic recomputation of angiography-derived Quantitative Flow Ratio (QFR), and validation against invasive fractional flow reserve (FFR). The system integrates state-of-the-art computer vision architectures with physically principled 1D hemodynamic modeling, enabling rapid, wire-free computation of coronary physiology and in-silico PCI planning. Evaluation on a cohort of 100 vessels demonstrated strong agreement with invasive FFR and high diagnostic performance, while operating fully automatically in the majority of cases (Kopanitsa et al., 9 Dec 2025).

1. System Architecture and Deep Learning Components

AngioAI-QFR is structured as a modular pipeline, with two primary computer vision modules forming the foundation:

A. Stenosis Detection

• The pipeline employs a YOLOv8m one-stage object detector with a CSPDarkNet backbone for lesion localization, supporting objectness, bounding-box regression, and class-probability branches.
• Training utilized 9,000 cine angiography frames with bounding box annotations curated by at least two interventional cardiologists for consensus.
• Data augmentation included random rotations (±15°), flips, intensity normalization, contrast-limited adaptive histogram equalization, and gamma correction.
• Loss is composed as $L_{det} = L_{cls} + L_{obj} + L_{bbox}$ with categorical cross-entropy for classification, binary cross-entropy for objectness, and CIoU/GIoU for box regression.
• Performance on held-out frames achieved precision of 0.966 and mAP@IoU 0.50 of 0.973; mAP@IoU [0.50–0.95] was 0.712.

B. Lumen Segmentation

• Lumen masks are produced by a DeepLabV3+ semantic segmentation network with a ResNet-50 backbone.
• Annotation included 250 pixel-wise annotated frames (train/val/test split: 175/50/25, no near-duplicate leakage).
• The loss function is pixel-wise cross-entropy ($L_{CE}$) optionally regularized by Dice loss for class imbalance.
• Post-processing comprises morphological cleanup (hole filling and small-component removal) and skeletonization for centerline extraction.
• On the test set, the intersection-over-union (IoU) was 0.643 and Dice similarity 0.781.

2. Anatomical Feature Extraction: Centerline and Diameter

Following segmentation, vessel morphology is quantified:

• Skeletonization (e.g., via Zhang–Suen thinning) reduces the lumen mask $M(x,y)$ to a 1D centerline $S(x,y)$.
• Local vessel radii $r_i$ are determined using the Euclidean distance transform $$D(x,y)= \min_{(u,v)\notin M} \sqrt{(x-u)^2 + (y-v)^2}$$, defining diameters $d_i = 2r_i$ at each centerline pixel.
• Skeleton pixels are sorted geodesically to yield $x \in [0, L]$ for curvilinear measurements along the vessel's main branch.

3. Functional Profiling: Relative Flow Capacity and QFR Calculation

A. Relative Flow Capacity (RFC)

• RFC at each position $x$ is defined by $RFC(x) = \left(\frac{d(x)}{d_{\rm ref}}\right)^4$, where $d_{\rm ref}$ is the proximal reference diameter, reflecting Hagen–Poiseuille scaling ($Q \propto r^4$).
• RFC is sampled at 1 mm intervals and discriminates focal (sharp, localized nadir) from diffuse capacity loss.

B. Angiography-QFR Model

• The vessel is discretized into $N$ segments of length $\Delta x \approx 1$ mm; for segment $n$, area is $A_n = \pi r_n^2$.
• Resting flow $Q_{rest}$ is derived from contrast transit time and cross-sectional area: $V_{rest} = L/(t_{dist}-t_{prox})$, $Q_{rest}=V_{rest}\cdot A_{ref}$; the surrogate hyperaemic flow $Q_{hyp}=k Q_{rest}$ uses a scaling factor $k$ calibrated offline.
• Segmentwise pressure loss combines viscous (Poiseuille) and local (Bernoulli) terms:
  • $$\Delta P_{{\rm visc},n} = \frac{8\mu Q_{hyp} \Delta x}{\pi r_n^4}$$
  • $$\Delta P_{{\rm loc},n} = K_n \frac{\rho Q_{hyp}^2}{2A_n^2}$$
  • Summed to total drop $$\Delta P_{tot} = \sum_{n=1}^N (\Delta P_{{\rm visc},n} + \Delta P_{{\rm loc},n})$$
• QFR is computed as $QFR = \frac{P_{dist}}{P_{prox}}$, where $P_{dist} = P_{prox} - \Delta P_{tot}$.

4. Virtual PCI Simulation and In-Silico Stenting

Virtual stenting in AngioAI-QFR enables prediction of post-PCI physiology:

• Stent landing zones ($x_{prox}$, $x_{dist}$) are selected by the user on the cine image or RFC profile.
• A target diameter curve $d_{tgt}(x)$ is defined as a smooth interpolation from proximal to distal reference diameters, constrained by stent size.
• A blended post-PCI profile $$d_{post}(x) = \alpha(x)d_{tgt}(x) + (1-\alpha(x))d(x)$$ is computed, with $\alpha(x)$ ramping from 0 (edges) to 1 (center).
• Post-intervention RFC and QFR are recalculated.
• In the 100-vessel cohort, median predicted QFR gain was +0.07 [$IQR$ 0.04–0.12] for focal disease ($\leq$20 mm RFC nadir) and +0.03 [$IQR$ 0.01–0.06] for diffuse disease ($>$20 mm).

5. Performance Evaluation Against Invasive FFR

Comparison with FFR in 100 consecutive vessels yields the following results:

Metric	Value	95% CI / IQR
Pearson correlation ($r$)	0.89	0.84 – 0.93
Mean absolute error (MAE)	0.045	—
Root mean square error	0.069	—
Bland–Altman bias	–0.008	–0.142 to 0.125
AUROC (FFR ≤ 0.80)	0.93	0.88 – 0.97
Sensitivity	0.88	—
Specificity	0.86	—
PPV	0.80	—
NPV	0.91	—
Accuracy	0.87	—
LAD AUROC	0.94	—
RCA AUROC	0.92	—
LCx AUROC	0.90	—

These results demonstrate strong agreement with invasive FFR. The AUROC of 0.93 at the FFR ≤ 0.80 threshold indicates high diagnostic discrimination. Vessel-specific AUROC suggests consistent performance across LAD, RCA, and LCx territories.

6. Workflow Characteristics, Automation, and Limitations

• Full automation was achieved in 93% of vessels, with 7% requiring minor ROI adjustment.
• Median total pipeline runtime was 41 s (IQR 31–58), with detection and segmentation averaging 12 s and physiology computation 7 s.
• Primary limitations include reduced segmentation and physiological accuracy in images with severe overlap or foreshortening, potential impact of bifurcations within the stented segment due to unmodeled side-branch losses, and the single-center, retrospective design of the validation cohort. Multicenter validation is pending.

7. Integration and Clinical Significance

AngioAI-QFR presents an integrated approach that unifies high-precision computer vision (YOLOv8m for stenosis detection, DeepLabV3+ for semantic segmentation) with classic 1D hemodynamics modeling for the real-time, wire-free estimation of QFR and virtual stent deployment. The pipeline’s demonstrated agreement with invasive FFR (r = 0.89, MAE = 0.045), strong overall diagnostic accuracy (AUROC = 0.93), high rate of fully automatic completion (93%), and sub-minute runtime suggest practical utility for physiology-guided PCI planning. RFC profiling further distinguishes focal from diffuse disease, enabling more granular lesion assessment and prediction of QFR gain with virtual stenting (Kopanitsa et al., 9 Dec 2025). A plausible implication is enhanced standardization and efficiency in the catheterization laboratory, pending further multicenter validation.