Segmentação e classificação automática de calcificação da artéria coronária em imagem de tomografia computadorizada utilizando aprendizado profundo

Abstract

One of the indicators of possible future occurrences of cardiovascular diseases is the amount of coronary arteries calcification. Recently, approaches using new technologies, such as deep learning, have helped to identify these indicators. This work proposes a method for segmenting and classifying coronary artery calcification. The segmentation task has 4 steps that are: extraction of the region of interest using U-Net with batch normalization after convolution layers, segmentation of calcifications, removal of false positives using U-Net with EfficientNetB0, and the classification step which was carried out based on the region where the calcification was segmented. The method uses a histogram matching technique as preprocessing, in order to increase the contrast between tissue and calcification and normalize the different types of exams. Tests were carried out between different architectures where the best approach in the segmentation task achieved an F1-Score rating of 96.9%, accuracy of 98.5% and sensitivity of 97.2% on the OrcaScore calcification dataset. In the classification task, the method achieved an Intraclass Correlation of 99.0% for all classes, and of 95.9%, 95.8% and 97.3% for the LAD, LCX and RCA classes, respectively, in addition to a weighted Kappa of 1.