Convolutional neural networks (CNN) are powerful tools for learning representations from images. They are being used in a large range of applications, being the state-of-the art in many computer vision tasks. In this work, we study the brain tumor segmentation problem using CNNs and the publicly available BraTS dataset. One of the key factors for this task is which training scheme is used since it should deal with memory constraints and should alleviate the high-imbalance nature between healthy and lesion tissue in the brain.

Thus, the purpose of this project is to propose a comparison between several training schemes and extensively analyze and evaluate them in terms of the dice score. We evaluate densetraining against patch-sampling, and particularly, xed-rule against adaptive sampling scheme. Furthermore, variants and modications of the existing training schemes have been proposed in order to enhance their performance. Finally, several loss functions for each training scheme have been analyzed.