Combining loss functions for deep learning bladder segmentation on dynamic MRI - Archive ouverte HAL Accéder directement au contenu
Proceedings/Recueil Des Communications IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) Année : 2021

Combining loss functions for deep learning bladder segmentation on dynamic MRI

Résumé

Pelvic floor disorders mainly affect women and turn to be a public health issue although their pathophysiology is still poorly understood. As the main concerned organs undergo motions and deformations that are specific to the pathology, dynamic MRI is a gold standard for radiologists. Unclear organ boundaries, inter-subject variability of organ shapes and pathological deformities make segmentation difficult to perform. To develop imaging software aimed at pathologies gradation, the accuracy of the segmentation of organ boundaries on MRI is an essential criterion. Automatic methods are not yet accurate enough to replace the mandatory manual segmentation step. Automatic segmentation methods using fully convolutional neural network (FCN) have been developed, but the loss functions used for their training were generally not sensitive enough for the detection of organ boundaries. We propose a loss function dedicated to organ boundary detection to enhance training and therefore improve results accuracy. The method was evaluated on bladder segmentation from dynamic 2D MRI, using a baseline U-Net architecture. We demonstrated that our network trained with a combination of a contour-based loss function and the widely used Dice loss outperformed existing methods and networks trained with the Dice loss alone.
Fichier non déposé

Dates et versions

hal-03327830 , version 1 (27-08-2021)

Identifiants

Citer

Marc-Adrien Hostin, Augustin C. Ogier, Nicolas Pirró, Marc-Emmanuel Bellemare. Combining loss functions for deep learning bladder segmentation on dynamic MRI. IEEE EMBS International Conference on Biomedical and Health Informatics (BHI), IEEE, pp.1-4, 2021, ⟨10.1109/BHI50953.2021.9508559⟩. ⟨hal-03327830⟩
45 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More