A recurrent neural network for computational physiology and its application to intracoronary image-derived fractional flow reserve

• Published in: Engineering applications of artificial intelligence Vol. 146; p. 110309
• Main Authors: Zheng, Sun, Wenbin, Jiao, Shuyan, Wang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.04.2025
• Subjects: Computational physiology; Coronary artery; Fractional flow reserve; Intracoronary imaging; Physical knowledge discovery; Recurrent neural network; Physical knowledge discovery; Recurrent neural network; Computational physiology; Fractional flow reserve; Intracoronary imaging; Coronary artery
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2025.110309

Traditional data-driven neural networks struggle to extract physical properties from complex visual data, particularly in medical imaging. An example is calculating fractional flow reserve (FFR) from coronary artery images, which involves the discovery of physical knowledge from visual data. This study proposes a novel deep learning framework to calculates the distribution of FFR along coronary vessels from intracoronary images, integrating both geometric features and blood flow dynamics. Existing FFR estimation methods rely on computationally intensive three-dimensional reconstructions, which may lack accuracy in estimating FFR distributions across vascular segments. A multi-input, multi-output deep bidirectional recurrent neural network is introduced, directly estimating FFR values from intracoronary images by combining vessel geometry with physical knowledge of blood flow, without the need for three-dimensional reconstructions. The network model demonstrates strong performance on synthetic and clinical datasets, with correlations of 0.91 and 0.85 to computational fluid dynamics simulations for synthetic and clinical data, respectively. It also achieves area under the receiver operating characteristic curve values of 0.97 and 0.94 for ischemic lesion detection. This framework enhances FFR estimation and ischemic lesion identification accuracy while reducing computational costs compared to three-dimensional reconstruction methods. The model shows promise for clinical applications in coronary artery diseases and could be integrated into clinical decision support systems to improve diagnostic workflows.