Generating Perfusion Maps from Multiphase Coronary CTA for Myocardial Ischemia Assessment
Abstract
Purpose
To develop an approach for deriving myocardial perfusion maps directly from multi-phase coronary CT angiography (CTA), eliminating the need for dedicated cardiac CT perfusion (CTP) protocols.
Methods
Ten pigs were examined in the supine position, yielding a total of 62 examinations acquired at weekly intervals. Each examination included a non-contrast CT (NCCT) followed by three-phase CTA (mCTA). The first phase consisted of a standard ECG- and contrast arrival–triggered coronary CTA, followed by two delayed phases acquired automatically after 6–7 seconds and 12–14 seconds. Time-resolved CTP images (15 volumes; temporal resolution 1.6–2 s) served as the reference standard. All mCTA and CTP images were deformably registered to the NCCT and reformatted into short-axis views. The myocardium was segmented, and perfusion maps were generated using model-based deconvolution from both NCCT + mCTA and CTP data as well. Due to the limited temporal resolution of mCTA, initial perfusion estimates were sensitive to noise and artifacts. To mitigate this, a multi-input conditional generative adversarial network (cGAN) was trained to refine mCTA-derived perfusion maps by using the initial maps and NCCT + mCTA images as input. CTP-derived perfusion maps served as the ground truth. Model accuracy was evaluated on a held-out 20% test set, at the subject level to prevent data leakage, using the standard 17-segment myocardial model.
Results
Initial perfusion maps derived from mCTA demonstrated a segment-wise mean absolute error (MAE) of 21.3 mL/100 g/min relative to the CTP maps. Refinement using the proposed cGAN reduced the MAE to 15.7 mL/100 g/min, corresponding to a 26.2% reduction.
Conclusion
This study demonstrates the feasibility of using multi-phase coronary CTA to generate perfusion maps, providing a more accessible approach with reduced radiation exposure. Future work will focus on enlarging the dataset by examining additional pigs to improve quantitative accuracy.