Classification of Coronary Calcifications using Deep Learning Approach: A Feasibility Study with the orCaScore Database

Abstract

This document presents a deep learning method for the automatic detection and classification of calcified lesions in coronary arteries using Convolutional Neural Networks (CNN) based on the AlexNet architecture. The demonstration of this proposal uses orCaScore, a standardized and labeled database of low-dose radiation computed tomography (CT) images of the heart. The methodology division was designed starting with the region of interest (ROI) extraction for consecutively utilizing a patch-based approach. We tested this approach in 7,386 patches and achieved an accuracy of 67%, sensitivity of 100%, precision of 67% and specificity of 75%. Our technique aims to reinforce the detection and quantification of coronary calcified lesions, enabling accurate diagnosis and treatment of cardiovascular diseases.