Encoder-Decoder Network for Brain Tumor Segmentation on Multi-sequence MRI

摘要

In this paper we describe our approach based on convolutional neural networks for medical image segmentation in a context of the BraTS 2019 challenge. We use the conventional encoder-decoder architecture enhanced with residual blocks, as well as spatial and channel squeeze & excitation modules. The present paper describes the general pipeline including the data pre-processing, the choices regarding the model architecture, the training procedure and the chosen data augmentation techniques. Our final results in the BraTS 2019 segmentation challenge are Dice scores equal to 0.76, 0.87 and 0.80 for enhanced tumor, whole tumor and tumor core sub-regions, respectively.