Machine learning in cardiovascular flows modeling: Predicting arterial blood pressure from non-invasive 4D flow MRI data using physics-informed neural networks (1905.04817v2)

Abstract: Advances in computational science offer a principled pipeline for predictive modeling of cardiovascular flows and aspire to provide a valuable tool for monitoring, diagnostics and surgical planning. Such models can be nowadays deployed on large patient-specific topologies of systemic arterial networks and return detailed predictions on flow patterns, wall shear stresses, and pulse wave propagation. However, their success heavily relies on tedious pre-processing and calibration procedures that typically induce a significant computational cost, thus hampering their clinical applicability. In this work we put forth a machine learning framework that enables the seamless synthesis of non-invasive in-vivo measurement techniques and computational flow dynamics models derived from first physical principles. We illustrate this new paradigm by showing how one-dimensional models of pulsatile flow can be used to constrain the output of deep neural networks such that their predictions satisfy the conservation of mass and momentum principles. Once trained on noisy and scattered clinical data of flow and wall displacement, these networks can return physically consistent predictions for velocity, pressure and wall displacement pulse wave propagation, all without the need to employ conventional simulators. A simple post-processing of these outputs can also provide a cheap and effective way for estimating Windkessel model parameters that are required for the calibration of traditional computational models. The effectiveness of the proposed techniques is demonstrated through a series of prototype benchmarks, as well as a realistic clinical case involving in-vivo measurements near the aorta/carotid bifurcation of a healthy human subject.

• The paper demonstrates that integrating physics-informed neural networks with 4D flow MRI accurately predicts arterial blood pressure non-invasively.
• The methodology leverages conservation laws and adaptive grid search to bypass complex simulation setups and mesh generation.
• Numerical results validate the approach against traditional solvers and indicate potential for real-time, patient-specific clinical diagnostics.

An Overview of Machine Learning in Cardiovascular Flows Modeling with Physics-Informed Neural Networks

The paper proposes a new methodology integrating machine learning techniques with computational modeling of cardiovascular flows. It demonstrates the application of physics-informed neural networks (PINNs) to predict arterial blood pressures using non-invasive 4D flow MRI data, challenging the conventional methods requiring invasive measurements and computationally expensive simulations.

Key Contributions

The authors focus on the integration of physics-based models with clinical data using PINNs, providing a robust framework that bypasses the necessity for complex mesh generation and extensive parameter tuning in traditional simulations. The method hinges on constraining neural networks with the fundamental conservation laws of mass and momentum, using one-dimensional models of pulsatile flow. Upon training with sparse and noisy flow data, the networks yield predictions for pressure, velocity, and wall displacement, critical parameters in characterizing cardiovascular health.

Notably, the paper highlights this work as the first application of PINNs to graphs and network topologies, addressing the physics across disjoint arterial domains using boundary conditions at interfaces like bifurcations. The authors utilize non-dimensionalization and normalization strategies to mitigate computational challenges such as vanishing gradients during training.

Numerical Results and Case Studies

The efficacy of the proposed methodology is examined through several synthetic and real-world case studies. The authors demonstrate significant agreement between the predictions of PINNs and benchmarks derived from a Discontinuous Galerkin solver. This includes a realistic simulation involving measurements from a human subject's aorta/carotid bifurcation.

Further, the authors use the dictionary from their neural network model to calibrate Windkessel model parameters. They propose an efficient method using adaptive grid search, emphasizing the practicality of PINNs in inferring unmeasured or challenging-to-measure parameters in clinical assessments.

Implications and Future Scope

The integration of PINNs with cardiovascular modeling holds substantial implications for clinical practice, potentially enabling non-invasive diagnostics and patient-specific surgical planning. The significant reduction in computational overhead associated with pre-processing and parameter tuning underscores the clinical viability of the proposed method, paving the way for real-time assessments directly from MRI data.

The work suggests future developments in AI-driven clinical tools, focusing on enhancing prediction accuracy, adapting to more complex tree-like arterial networks, and leveraging transfer learning to reduce computational time. Furthermore, addressing the remaining discrepancy in predictions, particularly concerning the smaller vascular networks and incorporating robust uncertainty quantification frameworks within PINNs, forms crucial steps ahead.

Conclusion

This paper exemplifies the power and potential of integrating machine learning, particularly physics-informed approaches, in predictive modeling of biological flows. The proposed methodology not only innovates upon current practices by making them more efficient and less invasive but also opens avenues for broader applications in computational physiology informed by large-scale clinical data.