3D Large Kernel Anisotropic Network for Brain Tumor Segmentation

Abstract

Brain tumor segmentation in magnetic resonance images is a key step for brain cancer diagnosis and clinical treatment. Recently, deep convolutional neural network (DNN) based models have become a popular and effective choice due to their learning capability with a large amount of parameters. However, in traditional 3D DNN models, the valid receptive fields are not large enough for global details from the objective and the large amount of parameters are easy to cause high computational cost and model overfitting. In order to address these problems, we propose a 3D large kernel anisotropic network. In our model, the large kernels in the decoders ensure the valid receptive field is large enough and the anisotropic convolutional blocks in the encoders simulate the traditional isotropic ones with fewer parameters. Our proposed model is evaluated on datasets from the MICCAI BRATS 17 challenge and outperforms several popular 3D DNN architectures.