Automatic CAC Voxel Classification with Multi-scale CNN Architecture

Coronary Artery Calcification (CAC) score is one of the most important measures in determining the degree of cardiovascular disease. It is time-consuming to do this manually or semi-automatically, so automatic CAC scoring methods are being studied. Most methods classify the calcified pixels(2D) or voxels(2.5D or 3D) and calculate the CAC score. We present a new automatic CAC voxel classification model with multi-scale CNN architecture which can reflect the advantages of large receptive CNN and small receptive CNN. This study used a cardiac CT dataset of 98 patients from Asan Medical Center in South Korea. The dataset consisted of a cardiac CT DICOM raw data and a CAC label data annotated by a cardiac radiologist. A total of 10,000 voxels were selected for each calcified artery(LM, CX, LAD, RCA) and background(BG), so a total of 50,000 voxels were used in training and testing. Our proposed model showed an accuracy of 89.58% with cardiac CT dataset of Asan Medical Center. The performance of our model is generally better when compared to other automatic CAC classification models.