PET Myocardial Flow Reserve Estimation from 4D-Coronary-CT using Deep Neural Network

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Computational Cardiology

Background:

• Myocardial flow reserve (MFR) is vital for assessing myocardial ischemic disease severity.
• Current NH3-positron emission tomography (NH3-PET) for MFR measurement is invasive and time-consuming.
• Non-invasive MFR estimation methods are needed to reduce patient and clinician burden.

Purpose of the Study:

• To develop and evaluate non-invasive methods for estimating MFR using coronary computed tomography angiography (CCTA) image processing.
• To compare the performance of MFR estimation using coronary flow index (CFI) with multiple linear regression (MLR) against vertical-time (VT) imaging with deep learning (FCN and CNN).

Main Methods:

• Two MFR estimation methods were proposed: 1) CFI with MLR, and 2) VT image analysis using fully connected networks (FCN) and convolutional neural networks (CNN).
• Performance was evaluated by comparing estimated MFR with observed values using the Pearson correlation coefficient for different coronary arteries (LAD, LCX, RCA).

Main Results:

• The MLR method showed Pearson correlation coefficients of LAD=0.31, LCX=0.19, and RCA=-0.21.
• The FCN method yielded correlations of LAD=0.19, LCX=-0.37, and RCA=0.23.
• The CNN method demonstrated superior performance with correlations of LAD=0.34, LCX=0.44, and RCA=0.37, outperforming the CFI-based method.

Conclusions:

• The VT image-based deep learning methods, particularly CNN, show superior performance for MFR estimation compared to the CFI-based MLR approach.
• Estimating MFR using VT images derived from CCTA is a promising, non-invasive alternative to NH3-PET.
• This approach holds significant clinical relevance for early diagnosis and management of cardiovascular diseases.