Constraint-Aware Learning for Fractional Flow Reserve Pullback Curve Estimation From Invasive Coronary Imaging

Area of Science:

• Cardiovascular Imaging and Intervention
• Medical Artificial Intelligence
• Computational Physiology

Background:

• Fractional flow reserve (FFR) pullback curve estimation from invasive coronary imaging is crucial for guiding coronary interventions.
• Current machine/deep learning methods for FFR estimation lack robust integration of geometric associations and physics principles.
• Accurate FFR assessment is vital for optimizing treatment strategies in patients with coronary artery disease.

Purpose of the Study:

• To develop and validate a novel constraint-aware learning framework for improved FFR pullback curve estimation.
• To incorporate geometrical and physical constraints to enhance the accuracy of FFR prediction along the coronary artery centerline.
• To leverage synthetic data and domain adaptation techniques to reduce clinical data dependency and bridge the synthetic-to-real data gap.

Main Methods:

• A constraint-aware learning framework integrating geometrical and physical constraints for FFR estimation.
• Utilization of synthetic data for model training to minimize clinical data collection costs.
• Implementation of a diffusion-driven test-time data adaptation method to align synthetic and real-world data distributions.
• Validation using both synthetic and a real-world dataset of 382 patients across three imaging modalities.

Main Results:

• The proposed method demonstrated superior performance in FFR estimation for stenotic coronary arteries compared to existing machine/deep learning and computational fluid dynamics models.
• High agreement and correlation were observed between the predicted FFR values and invasively measured FFR.
• The plausibility of FFR predictions along the coronary artery centerline was validated.

Conclusions:

• The constraint-aware learning framework offers a significant advancement in FFR pullback curve estimation from invasive coronary imaging.
• The integration of geometric and physics knowledge, coupled with data augmentation strategies, enhances predictive accuracy and clinical applicability.
• This approach holds promise for improving intraoperative guidance and patient outcomes in coronary interventions.