E-Locus - Institutional Repository of the University of Crete - GUDU: Geometrically-constrained Ultrasound Data augmentation in U-Net for echocardiography semantic segmentation

Home GUDU: Geometrically-constrained Ultrasound Data augmentation in U-Net for echocardiography semantic segmentation

Results - Details

[Add to Basket]

Identifier

000447242

Title

GUDU: Geometrically-constrained Ultrasound Data augmentation in U-Net for echocardiography semantic segmentation

Alternative Title

GUDU: γεωμετρικά προσδιορισμένη αύξηση δεδομένων για την σημασιολογική τμηματοποίηση εικόνων υπερήχου καρδιάς με χρήση συνελικτικών νευρωνικών δικτύων

Author

Σφακιανάκης, Χριστόφορος Γ.

Thesis advisor

Τζιρίτας, Γεώργιος

Reviewer

Αργυρός, Αντώνιος
Κομοντάκης, Νικόλαος

Abstract

Echocardiography is a very important medical examination that helps in the computation of critical heart functions. Boundary identification, segmentation and estimation of the volume of key parts of the heart, especially the left ventricle, is an important but difficult and time-consuming process, even for the most experienced cardiologists, due to shadows and speckle noise that characterize ultrasound images. In recent years, research has focused on the automatic segmentation of heart through artificial intelligence techniques and especially with the use of deep learning. Our work is part of this research framework. We implemented a neural network based on U-Net and trained it, using a large public dataset of cardiac ultrasound images (CAMUS dataset), to extract the areas of the left ventricle, myocardium and left atrium. In order to optimize the training process, we have developed a data augmentation method based on the medical practice in echocardiography. The evaluation of our method by the independent platform of the public competition CAMUS, showed an overall improvement in the segmentation accuracy but also in the estimation of the volume and the ejection fraction of the left ventricle. Specifically using the metric Dice for geometric metrics, the performance of our method for the epicardium reached 0.956 for the end-diastolic phase and 0.950 for the end-systolic phase. For the clinical metrics of the left ventricle volume, the Pearson correlation coefficient was used where our method gave 0.973, 0.974, 0.871 for the end-diastolic, end-systolic phase and ejection fraction respectively.

Language

English

Subject

Deep learning

Echocardiography

Image segmentation

Medical imaging

Ιατρική απεικόνιση

Νευρωνικά δίκτυα

Τμηματοποίηση

Υπερυχοκαρδιογραφία

Issue date