Asymmetric U-Net for Brain Tumor Segmentation: Transfer to an independent database

Abstract

An automatic and accurate brain tumor segmentation software for magnetic resonance imaging is crucial for clinical assessment, follow-up, and subsequent gliomas treatment. Convolutional Neural Networks (CNN) is the state-of-the-art in this task. One of the fundamental challenges for the inclusion of CNN's into clinical practice is the networks' ability to generalize their performance on a different dataset, other than the one in which the model was trained. Most of the proposed methods only evaluate their models on public databases and do not test them in real clinical images. We present a 3D Asymmetric U-Net for brain tumor segmentation from MRI images in patients with glioma. Our model has been trained on the BraTS 2020 public database. Besides, our model performance was evaluated on an independent cohort of 12 patients from the Bretonneau Hospital.