PUNCH: Physics-informed Uncertainty-aware Network for Coronary Hemodynamics

More than 10 million coronary angiograms are performed globally each year, providing a gold standard for detecting obstructive coronary artery disease. Yet, no obstructive lesions are identified in 70% of patients evaluated for ischemic heart disease. Up to half of these patients have undiagnosed, life-limiting coronary microvascular dysfunction (CMD), which remains under-detected due to the limited availability of invasive tools required to measure coronary flow reserve (CFR). Here, we introduce PUNCH, a non-invasive, uncertainty-aware framework for estimating CFR directly from standard coronary angiography. PUNCH integrates physics-informed neural networks with variational inference to infer coronary blood flow from first-principles models of contrast transport, without requiring ground-truth flow measurements or population-level training. The pipeline runs in approximately three minutes per patient on a single GPU. Validated on synthetic angiograms with controlled noise and imaging artifacts, as well as on clinical bolus thermodilution data from 20 patients, PUNCH demonstrates accurate and uncertainty-calibrated CFR estimation. This approach establishes a new paradigm for CMD diagnosis and illustrates how physics-informed inference can substantially expand the diagnostic utility of available clinical imaging.

💡 Research Summary

Coronary microvascular dysfunction (CMD) is a prevalent yet under‑detected cause of myocardial ischemia, traditionally diagnosed by invasive measurements of coronary flow reserve (CFR) using Doppler wires or thermodilution. Because these procedures are costly, risky, and require specialized equipment, the majority of patients undergoing coronary angiography receive no functional assessment. In this work the authors introduce PUNCH, a non‑invasive, uncertainty‑aware framework that estimates CFR directly from routine coronary angiograms. The method first extracts a vessel centerline from fluoroscopic sequences and builds a spatio‑temporal intensity map (kymograph) that reflects the concentration of contrast agent I(s,t) along the artery. Assuming that contrast transport can be described by a one‑dimensional advection‑diffusion equation, the authors embed this physics into a physics‑informed neural network (PINN). Three subnetworks predict (i) contrast intensity, (ii) axial blood velocity u(s), and (iii) an effective dispersion coefficient D(s,t). All subnetworks share a latent variable z, modeled with a Gaussian variational posterior qφ(z) and regularized by a Kullback‑Leibler (KL) divergence to a standard prior. The latent variable captures unobserved sources of uncertainty such as image noise, motion artifacts, and variability in contrast injection.

Training minimizes a composite loss L = L_data + L_PDE + λ_KL·KL, where L_data enforces fidelity to the observed kymograph, L_PDE penalizes violations of the advection‑diffusion PDE, and the KL term prevents over‑fitting while allowing the posterior to expand in low‑quality regions. Automatic differentiation supplies the required spatial and temporal derivatives. The network architecture is deliberately lightweight: the intensity predictor uses three hidden layers of 64 units, while the velocity and dispersion predictors each use two hidden layers of 32 units, with sigmoid‑based output bounds to enforce physiologically plausible ranges. Rest and hyperemic states are modeled with separate subnetworks that share the same latent z, ensuring coherent propagation of uncertainty between the two physiological conditions.

During inference, the posterior over z is sampled N=100 times (Monte‑Carlo). For each sample the model yields velocity fields under rest and hyperemia, which are spatially averaged to compute a per‑sample CFR = ū_hyper / ū_rest. The final CFR estimate is reported as the mean of the samples together with a 95 % confidence interval, providing a calibrated measure of predictive uncertainty.

The authors validate PUNCH on two fronts. First, synthetic angiograms with controlled noise, motion, and incomplete contrast filling demonstrate that the method remains accurate and that the 95 % confidence intervals contain the ground‑truth CFR in >95 % of cases, confirming proper calibration. Second, a clinical cohort of 20 patients with invasively measured CFR (thermodilution) shows strong agreement: Spearman’s ρ = 0.86, Bland‑Altman mean bias = −0.03 ± 0.12, and concordance correlation coefficient = 0.84. Moreover, the model’s uncertainty intervals encompassed the invasive measurements in 96 % of subjects, indicating that the network does not over‑confidently predict in noisy or ambiguous cases. Computationally, the entire pipeline runs in roughly three minutes per patient on a single GPU, making it feasible for routine clinical use.

Limitations include the reliance on a one‑dimensional centerline approximation (which may break down at bifurcations), focus solely on the left anterior descending artery, and the assumption of a relatively constant contrast injection profile. Future work will extend the framework to three‑dimensional vessel geometries, incorporate multi‑view angiography, and integrate systemic physiological variables (blood pressure, heart rate) into the latent representation. Large‑scale, multi‑center studies are planned to further assess clinical impact.

In summary, PUNCH demonstrates that embedding first‑principles physics into deep learning, combined with variational inference, enables accurate, fast, and uncertainty‑aware estimation of coronary flow reserve from standard angiograms, opening a new, non‑invasive pathway for diagnosing coronary microvascular dysfunction.