ATSNet: An Attention-Based Tumor Segmentation Network

Science and technology has had a huge impact in the field of medicine leading to more accurate and preventive diagnosis, and treatment. Detecting brain tumors in early stages is essential for timely treatment of patients. Automatic segmentation of brain tumors is a challenging task as tumors vary in shapes and size. In this paper, we propose a fully automatic novel deep learning architecture for brain tumor segmentation named ATSNet. The network provides an end-to-end solution for feature extraction and brain tumor segmentation on Magnetic Resonance Images. Our proposed model uses an encoder-decoder architecture, employing residual modules for tackling gradient dispersion and uses skip connections for better feature map synthesis. The network utilizes attention gates (AG) to tackle the variability of brain tumors. Performance metrics such as dice score, precision, recall and intersection-over-union (IoU) have been used to evaluate and benchmark our model against those reported in literature. We have evaluated our model using the k-fold cross-validation approach. Our analysis also includes an ablation study on our model to identify important parts of the architecture by their effect on performance for optimizing the model.