Deep Learning Methods for MRI Brain Tumor Segmentation: a comparative study

Ikram Brahim (1) , Dominique Fourer (1) , Vincent Vigneron (1) , Hichem Maaref (1)

1 IBISC - Informatique, BioInformatique, Systèmes Complexes

Brain tumor segmentation from MRI is an important task in biomedical image processing that can help specialists to predict diseases and to improve their diagnoses. Nowadays, most of the state-of-the-art techniques are based on deep learning neural networks for which the choice of the best architecture remains an open question. Hence, this paper aims at providing answers through an intensive and comprehensive comparison between several promising neural network architectures. Our study leads us to three approaches which are respectively based on 2D U-Net, 3D U-Net and cascaded neural networks, that are compared together and with another unsupervised technique based on k-mean clustering. We also consider several enhancement techniques such as data augmentation, curriculum learning and an original boosting method based on majority voting. We achieve to improve the results of the baseline methods in terms of Dice score when the suitable combination of techniques is used.

Mots clés: MRI; curriculum learning; Brain tumor segmentation; deep learning

Domaines: Traitement du signal et de l'image [eess.SP]; Imagerie médicale; Intelligence artificielle [cs.AI]

Ikram Brahim, Dominique Fourer, Vincent Vigneron, Hichem Maaref. Deep Learning Methods for MRI Brain Tumor Segmentation: a comparative study. 9th IEEE International Conference on Image Processing Theory, Tools and Applications (IPTA 2019), Nov 2019, Istanbul, Turkey. ⟨10.1109/IPTA.2019.8936077⟩. ⟨hal-02296594⟩