3D Dense Dilated Hierarchical Architecture for Brain Tumor Segmentation

Abstract

In the medical domain, 3D convolutional neural network has accomplished excellent results in analysing brain MRI. Inspired by the dense connection, dilation, types of feature fusion methods and availability of essential standards of information presented in multi-modalities and hierarchical knowledge transfers, we present a unique solution for automatically segmenting high-grade glioblastoma, which is a critical type of brain tumor leading to a large number of deaths in the world. The improved survival rates during early detection of this tumor encourage us to provide our contribution using popular machine learning techniques, such as deep learning. In such contribution, the convolutional layer automatically handles the learning feature with a vital concept of receptive field. The deep network allows increased receptive field for learning excellent information. However, this method can be halted by limited memory, especially in building a deep hierarchical architecture. Our model provides a solution towards the path of building a deep, dense architecture by implementing the dilation layer with dense pattern. The receptive field can be increased exponentially and the excessive use of convolution layers can be reduced to reduce parameters and allow the efficient utilisation of GPU memory. Effectively combining local and global contextual information in our model helps in further understanding of different tumor types. We propose to segment the entire tumor hierarchically and use its information to segment its sub-regions, thereby enhancing the tumor and tumor core. The network trained and validated with Brats 2018 data set can achieve dice scores of 0.8480, 0.8574 and 0.8219. Results verify the utilisation of dense connectivity with mixed dilated and conventional convolutional layers. To learn the global contextual information, we initially increase and then decrease the dilation rated with dense connections to minimise the possible loss of information due to the successive dilated rates. Our experimental results verify our innovative work for utilising this type of connectivity in medical image segmentation.