Deep Learning-Based Classification of Invasive Coronary Angiographies with Different Patch-Generation Techniques

Medical imaging is one of the areas where computer-aided diagnosis could improve the efficiency of diagnosis in clinical settings. Cardiovascular artery disease (CAD) is diagnosed by invasive coronary angiography (ICA). This paper reports on performance analysis for binary classification of ICA images by grouping severity ranges and evaluates how performance is affected by the degree of lesions and the patch generation technique considered. An annotated dataset of ICA images was used, categorizing lesions into seven possible ranges: <20%, [20%, 50%), [50%, 70%), [70%, 90%), [90%, 98%], 99% and 100%. In this study, three pre-trained CNN architectures were trained using different categories of lesion severity as input, and their F-measures and accuracy were computed, achieving a performance above 90%.