Reaching intra-observer variability in 2-D echocardiographic image segmentation with a simple U-Net architecture
Résumé
Segmentation of the endocardial and epicardial contours of the left ventricle in echocardiography has been the subject of much research for years. Hitherto, only CLAS managed to achieve performance comparable to intra-observer variability by using a nontrivial loss function. We carried out an extensive study of nnUNet to investigate whether a simple U-Net architecture could have equivalent segmentation performance to intra-observer variability. Our study shows that the data augmentation in both training and inference, combined with a well-matched optimization scheme, is the key to achieve this goal. Consequently, a simple U-Net architecture can produce high-quality segmentation and accurate volume estimation.
Fichier principal
Reaching intra-observer variability in 2-D echocardiographic image segmentation with a simple U-Net architecture V2.pdf (356.72 Ko)
Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Licence : CC BY - Paternité
Licence : CC BY - Paternité