Unet3D with Multiple Atrous Convolutions Attention Block for Brain Tumor Segmentation

Brain tumor segmentation by computer computing is still an exciting challenge. UNet architecture has been widely used for medical image segmentation with several modifications. Attention blocks have been used to modify skip connections on the UNet architecture and result in improved performance. In this study, we propose the development of UNet for brain tumor image segmentation by modifying its contraction and expansion block by adding Attention, adding multiple atrous convolutions, and adding a residual pathway that we call Multiple Atrous convolutions Attention Block (MAAB). The expansion part is also added with the formation of pyramid features taken from each level to produce the final segmentation output. The architecture is trained using patches and batch 2 to save GPU memory usage. Online validation of the segmentation results from the BraTS 2021 validation dataset resulted in dice performance of 78.02, 80.73, and 89.07 for ET, TC, and WT. These results indicate that the proposed architecture is promising for further development.