A Deep Learning Approach for Brain Tumour Segmentation using Connected-UNets

Abstract

Gliomas are the most prevalent and deadly type of brain tumor, with an extremely low life expectancy in their most severe forms. Thus, a crucial step in enhancing the quality of life for cancer patients is treatment planning. Although Magnetic Resonance Imaging (MRI) is a popular imaging method for evaluating these tumors, the volume of data it generates makes it difficult to manually segment the images in a reasonable length of time, which restricts the use of precise quantitative assessments in clinical practise. The enormous spatial and structural heterogeneity among brain tumors makes automatic segmentation a difficult task, hence effective and automatic segmentation methods are needed. UNet and its variants are one of the most advanced models for medical image segmentation, and they performed well on MRI images. So, in this paper we designed an automatic brain segmentation approach using Connected-UNets, which connects two UNets using additional modified skip connections. To highlight the contextual information within the encoder-decoder network design, we integrate Atrous Spatial Pyramid Pooling (ASPP) in the two conventional UNets. We also apply the proposed architecture on the Attention UNet (AUNet) and the Residual UNet (ResUNet). To evaluate the proposed model BraTS 2017 dataset with HGG and LGG images are considered. The experimental results give Dice Coefficient for the three architectures Connected UNets, Connected AUNets, and Connected ResUNets as 90.12%,91.45%, and 92.2%, and IoU score as 86%, 88%, and 90% respectively.