Dilated Convolutions based 3D U-Net for Multi-Modal Brain Image Segmentation

Abstract

Several deep learning based medical image segmentation methods use U-Net architecture and its variants as a baseline model. This is because U-Net has been successfully applied to many other tasks. it was noticed that the U-Net-based models are unable to extract features for segmenting small masks or fine edges. To overcome this issue, we propose a new 3D U-Net-based model, baptized Y- Net. In this model, we make use of dilated convolution which has shown its effectiveness in grasping different features at different scales. This allows us to capture more information from small anatomical parts. Our model is assessed on MRbrains13 dataset for brain tissue segmentation task. Compared to the traditional UNet 3D, the obtained results show that the proposed model performs well, especially in segmenting white matter and grey matter tissues.