Right coronary artery labelling with point annotations
| Autor: | Ferrero Montes, Laura |
|---|---|
| Přispěvatelé: | Menkovski, Vlado, Oliván Bescós, Javier, Ozcelebi, Tanir |
| Rok vydání: | 2019 |
| Předmět: | |
| Zdroj: | Archivo Digital UPM Universidad Politécnica de Madrid |
| Popis: | Coronary Artery Disease (CAD) is a disorder that affects the coronary arteries and it is characterized by the narrowing or blockage of these vessels. Since they provide oxygen-rich blood to the cells of the cardiac muscle, CAD can produce an angina or heart attack. This disorder is diagnosed using angiography and this imaging technique is also used during the treatment of this disease so physicians can determine the exact position of the lesion in real-time. The pathological findings associated with CAD must be reported per arterial branch. In consequence, an automatic labelling of the different branches will be very useful for physicians and it will also be beneficial for the quantitative characterization of the coronary structure. Research in this field has been focused on the extraction of the centerlines of the arteries and the classification of each of them in one of the coronary artery tree divisions. A common strategy used to label the centerlines is to map them to a model that encodes the clinical knowledge but the main challenge of the arterial structure is the heterogeneity among patients that is difficult to include in such a model and the local similarity among the segments. Such approaches, however, depend on the quality of the extracted centerlines so they cannot solve directly the problem and no previous research has tried to identify the branches directly from the images. Deep learning has shown promising results in other tasks related with medical imaging but the main limitation of deep learning for medical data is the amount of information necessary to train the networks. In addition, the ground-truth data must be annotated by medical experts. To overcome this problem, this project proposes an algorithm to label the different segments of the right coronary tree by the regression of the start and end point of each branch. Therefore, the annotation process is more efficient. In addition, this proposal directly works with angiographies, thereby it does not need previous steps as segmentation or centerline extraction. This algorithm has a fully convolutional architecture and it includes the hierarchical structure of the data. The information from points that are connected in the coronary tree is included for the final prediction of the model. It was trained with 111 angiographies and evaluated with 23 images. The performance of the model was measured with the average Euclidean distance between the predicted coordinate of the point and the ground-truth. The error obtained in the location of the different points varies from 4 to 8% of the image. However, more images that cover a higher amount of the data heterogeneity would be helpful to improve the generalization of the model. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|