Deep learning for coronary artery disease severity classification

Abstract

Medical imaging evaluations are one of the fields where computed-aid diagnosis could improve the efficiency of diagnosis supporting physician decisions. Cardiovascular Artery Disease (CAD) is diagnosed using the gold standard, Invasive Coronary Angiography (ICA). In this work, performance analysis for binary classification of ICA images considering the severity ranges separately is reported, evaluating how performance is affected depending on the degree of lesions considered. For this purpose, an annotated dataset of ICA images was employed, which contains the ground truth, the location and the category of lesions into seven possible ranges: <20 %, [20 %, 49 %], [50 %, 69 %], [70 %, 89 %], [90 %, 98 %], 99 %, and 100 %. The ICA images were pre-processed, divided into patches and balanced by downsampling and data augmentation. In this study, four known pre-trained CNN architectures were trained using different categories of lesion degree as input, whose F-measures are computed. Results report that the F-measures showed a behavior dependent on the narrow presents of the image, being lesions with more than 50 % severity were better classified, achieving an F-measure of 75%.