A new Online Class-Weighting approach with Deep Neural Networks for image segmentation of Highly Unbalanced Glioblastoma Tumors - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2019

A new Online Class-Weighting approach with Deep Neural Networks for image segmentation of Highly Unbalanced Glioblastoma Tumors

Résumé

The most common problem among image segmentation methods is unbalanced data, where we find a class or a label of interest has the minority of data compared to other classes. This kind of problems makes Artificial Neural Networks, including Convolutional Neural Networks (CNNs), bias toward the more frequent label. Thus, training a CNNs model with such kind of data, will make predictions with low sensitivity, where the most important part in medical applications is to make the model more sensitive toward the lesion-class, i.e. tumoral regions. In this work, we propose a new Online Class-Weighting loss layer based on the Weighted Cross-Entropy function to address the problem of class imbalance. Then, to evaluate the impact of the proposed loss function, a special case study is done, where we applied our method for the segmentation of Glioblastoma brain tumors with both high-and low-grade. In this context, an efficient CNNs model called OcmNet is used. Our results are reported on BRATS-2018 dataset where we achieved the average Dice scores 0.87, 0.75, 0.73 for whole tumor, tumor core, and enhancing tumor respectively compared to the Dice score of radiologist that is in the range 74%-85%. Finally, the proposed Online Class-Weighting loss function with a CNNs model provides an accurate and reliable seg-mentation result for the whole brain in 22 seconds as inference time, and that make it suitable for adopting in research and as a part of different clinical settings.
Fichier principal
Vignette du fichier
paper_174.pdf (403.64 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Loading...

Dates et versions

hal-02172208 , version 1 (03-07-2019)

Identifiants

Citer

Mostefa Ben Naceur, Rostom Kachouri, Mohamed Akil, Rachida Saouli. A new Online Class-Weighting approach with Deep Neural Networks for image segmentation of Highly Unbalanced Glioblastoma Tumors. International Work-Conference on Artificial Neural Networks (IWANN-Advances in Computational Intelligence), Jun 2019, Gran Canaria, Spain. ⟨10.1007/978-3-030-20518-8_46⟩. ⟨hal-02172208⟩
81 Consultations
689 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More