Abstract
Objective . To provide a translation-oriented review of imaging-derived coronary fractional flow reserve (FFR) methods for functional assessment of coronary artery stenosis, with emphasis on computed tomography coronary angiography- and invasive coronary angiography-based workflows. Approach . We synthesize recent developments in physics-based computational fluid dynamics (CFD), reduced-order modeling, machine learning/deep learning (ML/DL), and physics-informed learning, including physics-informed neural networks (PINNs) and neural operators. The review is organized around clinical workflow requirements, computational feasibility, image-quality constraints, boundary-condition uncertainty, model validation, and deployment-oriented evaluation. Main results . Physics-based FFR methods provide mechanistic interpretability and spatially resolved hemodynamics, but remain sensitive to coronary geometry, physiological assumptions, and computational burden. ML/DL approaches markedly accelerate FFR estimation, but their reliability depends on representative data, calibration, and robustness to domain shift. Among emerging approaches, physics-informed learning currently offers a favorable methodological balance of computational speed, physical consistency, data efficiency, and spatially resolved output. Complementary coronary studies support this assessment: graph-based physics-informed learning has shown strong agreement with measured FFR across x-ray and CT data; a physics-informed operator preserved near-CFD/PINN diagnostic performance while reducing inference from hundreds of seconds to less than one second; and physics-informed pretraining improved a downstream clinical prediction task while producing pressure and FFR curves. The latter finding supports transferable physiological representation learning rather than direct validation of FFR accuracy against invasive pressure-wire measurements. Overall, these results support physics-informed methods as a promising next-generation research direction, although prospective multicenter validation remains necessary. Significance . Imaging-derived FFR is moving toward faster, more automated, and more physiologically constrained decision support. Among emerging computational paradigms, physics-informed models are particularly promising because they can retain the speed of learned inference while constraining predictions with governing equations and boundary conditions. Current comparative evidence supports prioritizing these methods for prospective clinical validation, with their apparent advantages tested against established CFD, data-driven ML, and invasive physiological references.
Funding Information
  • Michigan Technological University Health Research Institute
  • American Heart Association (25AIREA1377168)
  • National Institutes of Health, USA (1R15HL172198)

This publication has 67 references indexed in Scilit: