Non‐invasive fractional flow reserve estimation using deep learning on intermediate left anterior descending coronary artery lesion angiography images
Code: G-1054
Authors: Aydin Sadeghi * ℗
Schedule: Not Scheduled!
Tag: Biomedical Signal Processing
Download: Download Poster
Abstract:
Abstract
Introduction: Coronary artery disease (CAD) is a predominant global health challenge, with fractional flow reserve (FFR) widely recognized as the clinical gold standard for assessing the hemodynamic significance of coronary artery stenosis. Nevertheless, traditional FFR measurement is invasive, costly, and associated with procedural risks, thus limiting its routine clinical application. Recent advances in artificial intelligence (AI) present an opportunity to non-invasively estimate FFR directly from angiography images, potentially revolutionizing the diagnostic and therapeutic approach to CAD by combining both anatomical and physiological insights. Methods: This study introduces an AI-driven, end-to-end deep learning model for estimating FFR, targeting intermediate stenosis in the left anterior descending (LAD) artery (50-70% stenosis). Leveraging a dataset of 3,625 LAD angiography images from 41 patients, we evaluated nine pre-trained convolutional neural networks (CNNs) and identified DenseNet169 as the optimal model for image feature extraction and FFR classification. Rigorous data preprocessing steps—including resizing, normalization, and data augmentation—were applied to enhance model generalization. DenseNet169 was meticulously fine-tuned for maximum sensitivity, specificity, and classification accuracy, with performance metrics encompassing area under the curve (AUC), F1-score, precision, and recall. Results: The DenseNet169 model demonstrated exceptional diagnostic accuracy, achieving an AUC of 0.81, sensitivity of 0.86, specificity of 0.75, and overall accuracy of 81%. These results validate the model’s capacity to reliably distinguish functionally significant from non-significant coronary stenosis, closely mirroring the diagnostic accuracy of invasive FFR. This non-invasive approach offers a fast, reliable, and resource-efficient tool for accurate FFR estimation, streamlining the CAD diagnostic workflow and reducing dependency on invasive measurements. Conclusion: This AI-powered FFR estimation model represents a significant advancement in CAD diagnostics, providing a precise, non-invasive alternative that aligns with clinical standards while circumventing the limitations of traditional FFR. By integrating physiological assessment into routine catheterization lab workflows, this model has the potential to enhance decision-making, improve patient outcomes, and optimize healthcare resource allocation. Future studies with broader clinical validation could further substantiate its clinical utility, paving the way for AI-driven precision medicine to redefine CAD management and enhance the quality of cardiovascular care globally.
Keywords
Intermediate lesions, Coronary blood flow Fractional
