An Improved Framework Called Du++ Applied to Brain Tumor Segmentation

Abstract

We all know the advanced framework which is used to medical image processing is Unet, but it is struggling when it processes complex images. DenseNet is the state-of-the-art network, which has large parameters compared with Unet. Unet++ performs better on complex images than Unet. In this work, we proposes an novel network structure called Dense_Unet++(DU++), that can take advantage of feature fusion of the Unet++, reduces the DenseNet's parameters and further improves the segmentation accuracy. Our model is mainly implemented by combine Half Dense Unet(HDU)and Unet++. The long connections with different semantic levels do not achieve the effect of feature fusion, so our paper propose that built a series of bridges for different semantic levels within the DU++ and abandoned the original long connections. We apply this framework to brain tumor segmentation. In the end, our experiment achieved a promising result.