In a paper published in the journal PNAS Nexus, researchers have addressed the complexity of blood velocity and red blood cell (RBC) distribution in microcirculation capillaries. Existing imaging techniques cannot directly measure full 3D velocity and RBC concentration profiles, which is crucial for assessing physiological variables like wall shear stress (WSS) and cell-free layer (CFL).
Study: Microvascular Hemodynamics: Machine Learning Empowers Precision Predictions. Image credit: Rost9/Shutterstock
Traditional theoretical models and high-fidelity computational approaches must provide comprehensive 3D profiles across large spatial scales. The researchers introduced machine learning (ML) models based on artificial neural networks and convolution-based U-net models to overcome these limitations. These ML models accurately predict hemodynamic quantities in microvascular networks, significantly reducing prediction time while maintaining high accuracy. This work allows ML to enhance hemodynamic predictions in organ-scale microvascular networks.
Related Work
Previous studies underscore the vital role of the microvascular network in tissue function, emphasizing the importance of understanding blood flow dynamics and RBC distribution for metabolic exchange. However, obtaining complete 3D profiles of velocity and RBC concentration remains challenging due to the complex nature of microvascular flow and limitations in existing imaging techniques. Traditional theoretical and computational models lack scalability for providing detailed profiles across large spatial scales.
ML Techniques
The data utilized in this study originates from high-fidelity 3D simulations representing the flow of deformable RBC suspensions within physiologically realistic microvascular networks, mimicking in vivo conditions. These networks, denoted as vasculatures A and B, are complex geometric constructs with multiple vessels, bifurcations, and mergers, approximating tissue areas of significant size. In these simulations, blood, composed of RBCs and plasma, flows through the virtual vasculatures, with about 1,000 RBCs present in each at any given time over a simulated period of approximately 1.5 seconds, surpassing an average cardiac cycle.
The computational methodology employed for these simulations integrates finite volume, finite element, and immersed-boundary methods implemented within a coupled framework. It allows for an accurate representation of fluid dynamics and RBC behavior within the microvascular networks. Researchers modeled RBCs as deformable entities, each comprising hemoglobin enclosed by a membrane, exhibiting complex shapes and distributions resembling those observed in vivo.
The simulations provide comprehensive 3D profiles of fluid velocity and RBC concentration in every vessel of the microvascular network, which is crucial for understanding hemodynamics. However, extracting meaningful insights from these vast datasets poses significant computational challenges. To address this, researchers leverage ML techniques to actively develop models capable of accurately and efficiently predicting hemodynamic variables.
These ML models are trained using vasculature A data and tested on vasculature B to ensure generalizability. Separate models are constructed for different vascular components, including vessels, bifurcations, and mergers, considering each component's distinct flow dynamics and RBC distribution mechanisms2D and 3D models are developed, with the latter employing convolutional neural network-based U-net architecture to capture spatial variations in velocity and RBC concentration.
The ML models demonstrate remarkable accuracy in predicting hemodynamic profiles, significantly reducing computational time compared to traditional high-fidelity simulations. These ML-based approaches offer a promising avenue for advancing the understanding of tissue perfusion and vascular pathophysiology by enabling precise predictions of complex hemodynamics within large-scale microvascular networks.
ML Models for Hemodynamics
This study introduces ML models as a novel approach to predicting detailed hemodynamic parameters in microvascular networks. Traditional imaging techniques cannot directly measure blood velocity and RBC concentration profiles, which are essential for understanding hemodynamics. High-fidelity simulations offer such details but are computationally intensive. ML models, trained on simulation data, offer a faster and more efficient alternative. By employing regression-type artificial neural networks (ANN) and convolution-based U-net models, the study achieves accurate blood velocity and RBC concentration profile predictions with minimal mean squared error (MSE).
While high-fidelity models provide comprehensive data and reveal new physics, they demand significant computational resources and expertise. In contrast, ML models offer a trade-off between accuracy and computational efficiency. They are instrumental when specific hemodynamic variables, like WSS distribution, are fascinating. The reduced computational time and broader accessibility make ML models viable for understanding diseases like retinopathy, Alzheimer's, and dementia. Moreover, they can be adapted for in vivo imaging data, providing detailed 3D hemodynamic parameters.
Despite their promise, the current ML models have limitations. They assume cylindrical vessels with constant diameter and do not account for vessel curvature. Additionally, they lack physics-informed constraints and can only predict time-averaged hemodynamics. Future models could incorporate vessel geometry variations and physics-based constraints to address these limitations. Furthermore, extensions of ML models could explore applications in blood cell disorders, drug transport in diseased vasculature, and adaptations during vascular development and dysfunction.
In conclusion, the study presents pioneering ML models for predicting subcellular resolved capillary hemodynamics in organ-scale networks. These models offer promise for understanding various diseases and physiological processes, including blood disorders and vascular adaptations. Future research could refine the models to accommodate diverse vascular geometries, transient conditions, and abnormal physiological states. Moreover, applying ML in non-biological domains, such as fluid dynamics in porous media, holds potential for broader scientific and engineering applications.
Conclusion
In summary, the introduction of ML models represented a significant advancement in predicting detailed hemodynamic parameters within microvascular networks. These models offered a faster and more efficient alternative to traditional imaging techniques and high-fidelity simulations.
Despite their limitations, including assumptions about vessel geometry and lack of physics-informed constraints, ML models held promise for understanding various diseases and physiological processes. Future research could refine these models to accommodate diverse conditions and expand their applications beyond biological domains, offering broader scientific and engineering insights.
Machine Learning Models Predict Arsenic Contamination