Coronary artery segmentation using Transformer Neural Networks

Abstract

Coronary Artery Disease (CAD) is the leading cause of death in developed countries. It is a multi-factorial disease consisting of a plaque accumulation in the coronary vessels, causing ischemia or myocardial infarction. An early diagnosis is important to avoid fatal consequences. Nowadays, Computed Tomography (CT) images are used as a diagnostic tool as well as a technique for selecting appropriate therapies for CAD patients. In order to compute disease-related metrics, it is necessary to process these images, including the coronary artery delineation. In the last years, the use of Artificial Intelligence (AI) has grown exponentially in clinical environments, especially in time-consuming tasks, such as image segmentation. There exist plenty of AI algorithms that have proven good performance in these tasks, including Transformers Neural Networks. Hence, the main aim of this project was to develop a coronary artery segmentation algorithm using this approach and study its performance to evaluate its potential in clinical practice. The results showed segmentations with the coronary artery shape well-defined but with several stops in the segmentation of the main branches and a huge presence of artefacts. These could be solved by computing a longer training using an extended dataset in the future, allowing their implementation in the clinical field. As healthcare professionals would only be responsible for the validation of the segmentation, they could devote more time studying markers to enhance the diagnosis of patients and provide a more personalised treatment. The created algorithm may work nowadays as a support tool in semi-automated segmentation.