U-InceptAtt: U-Net-Like Architecture with Inception Module Encoder-Decoder and Attention Bottleneck for Brain Tumor Segmentation

Glioma is one of the most dangerous and frequent primary brain tumors, it can spread quickly in the other parts of the brain which makes it very aggressive and requires very fast radiotherapy. Thus, accurate and reliable tumor segmentation is a crucial step in diagnosis, choosing a course of treatment, and identifying risk factors. Manual methods proposed in this area showed their complexity since they take a long time and depend on many specialists’ interventions. To address these limitations, deep learning automatic-based methods have been proposed and shown great success in this field, especially the U-Net extension architectures. In this paper, and based on the U-Net structure, we built our own model using the inception module in both the encoder and decoder parts, besides applying an attention mechanism to the bottleneck layer. The proposed approach was evaluated on the BraTS 2020 dataset to segment the different types of tumors. The experiment results showed a good performance in terms of dice similarity coefficient by achieving 89.44%, 81.42%, and 71.83% for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET), respectively.