Semantic segmentation of brain tumor with nested residual attention networks

Brain tumors are one of the most serious brain diseases, which often result in a short life. However, in developing areas, medical resources are in shortage, which affect the diagnosis of brain tumors. With the development of computer science, many diseases can be diagnosed by telemedicine systems, which help physicians save much time and improve diagnostic accuracy. Therefore, we propose a semantic segmentation method for brain tumors based on nested residual attention networks. It can be deployed in social mx‘edia environment to work as a remote diagnosis system. The proposed method uses an improved residual attention block (RAB) as the basic unit. Based on the improved RAB, a nested RAB is designed to build the proposed method, which has better generalization. The proposed method includes an encoder part, a decoder part and skip connections. The encoder part learns multiple feature representations from inputs and the decoder part utilizes the learnt features to make segmentation predictions. In addition, high-level attention feature maps are exploited to induce low-level feature maps in skip connections to discard useless information. The proposed method is mainly validated on the dataset of Brain Tumor Segmentation challenge (BraTS) 2015. The proposed method achieves an average dice score of 0.87 (0.80, 0.72) for the whole tumor (core tumor, enhancing tumor) regions on BraTS 2015 dataset.