A Robust Agatston Score for Coronary Artery Calcium Scoring From Non-Ecg-Gated Ct with Different Reconstruction Kernel

A coronary artery calcium score (CACS) is a vital measure to screen individuals at risk for early coronary heart diseases. However, as the main evaluation system of CACS, the Agatston score computed from CT images via HU-thresholding may vary significantly even for the same individual as the protocol of image acquisition changes (e.g., reconstruction kernels). This may harm the compatibility of CACS, when evaluated in different health facilities at different times. To tackle this issue, we propose the robust Agatston score (RAS), wherein we predict the calcification level per pixel via deep learning, rather than directly thresholding the HU value from CT images, as we do for the classic Agatston score. In this way, we make the CACS more robust to the change of acquisition protocols, and let the comparison among CACS from various sources easier. Experimental results show that our method can improve the CACS level accuracy from $64.21\%$ to $95.78\%$. We will release the code after paper acceptance.