Super-resolution left ventricular flow and pressure mapping by Navier-Stokes-informed neural networks.

Area of Science:

• Cardiovascular imaging and fluid dynamics.
• Application of artificial intelligence in medical diagnostics.
• Biomedical engineering and computational physiology.

Background:

• Current vector flow mapping (VFM) echocardiography uses kinematic constraints but lacks flow physics information, limiting its ability to derive pressure maps or integrate data across time frames.
• Existing VFM models are unable to fully capture the complex dynamics within the left ventricle (LV), hindering comprehensive analysis.
• The absence of flow physics, forces, and accelerations in traditional VFM restricts its diagnostic potential.

Purpose of the Study:

• To develop an advanced vector flow mapping (VFM) technique, termed AI-VFM, leveraging artificial intelligence (AI) and physics-informed neural networks (PINNs).
• To integrate fundamental fluid dynamics principles, including mass conservation and momentum balance, into echocardiographic VFM for enhanced accuracy.
• To enable AI-VFM to recover complete flow and pressure fields within the left ventricle (LV) from standard echocardiographic data, including areas with limited or no Doppler input.

Main Methods:

• Developed AI-VFM using PINNs that encode mass conservation and momentum balance within the LV, along with no-slip boundary conditions at the endocardium.
• Utilized standard echocardiographic color-Doppler sequences as input for the AI-VFM model.
• Validated AI-VFM performance using physiological simulated LV data, comparing models informed by mass conservation alone versus those including momentum balance.

Main Results:

• AI-VFM successfully recovered complete LV flow and pressure fields, performing phase unwrapping and filling data gaps where color-Doppler information was absent.
• The AI-VFM model achieved temporal super-resolution, generating complete flow maps at time points lacking direct Doppler input.
• Incorporating momentum balance into the PINNs significantly improved AI-VFM accuracy and enabled temporal super-resolution, outperforming models using only mass conservation.

Conclusions:

• AI-VFM represents a significant advancement in echocardiographic flow mapping, providing accurate, physics-informed flow and pressure visualization within the LV.
• The AI-VFM approach demonstrates the potential for super-resolution flow mapping and improved diagnostic capabilities by integrating patient-specific flow physics.
• The AI-VFM framework is computationally efficient (15 min on standard GPUs) and adaptable for mapping other cardiovascular metrics, offering broad clinical potential.