Automated Brain Tumor Segmentation Using Attention gate Inception UNet with Guided Decoder

Abstract

Brain tumor segmentation technology is a crucial step for the detection and treatment of MRI brain tumors. Tumors can occur in various locations and can be of any size or form. The use of skip connections in MRI brain tumor segmentation approach based on U-Net architecture helps to incorporate lowlevel and high-level feature information and has recently gained popularity. By introducing an attention mechanism into the UNet architecture, the performance of local feature expression and medical image segmentation can be enhanced. In this paper, we present an innovative deep learning architecture called Attention gate Inception UNet with Guided Decoder for brain tumor segmentation. The backbone of the model is a popular segmentation method called U-Net architecture. While dealing with small-scale tumors, the U-Net network has low segmentation accuracy. Therefore several modifications are made, which results in the integration of attention gates and inception block together with a guided decoder. A sequence of attention gate modules are introduced to the skip connection, that focus on a selected part of an image while ignoring the others. The inception module used will help us to extract further characteristics at each layer. The proposed architecture has the ability of explicitly guiding each decoder layer’s learning process and it is supervised by using individual loss function, allowing them to produce efficient feature maps. The proposed model achieved a dice score of 0.9190, 0.9331, 0.8990 for whole tumor, tumor core and enhancing tumor respectively on Brain Tumor Segmentation Challenge (BraTS) 2019 dataset of High Grade Gliomas (HGG)