Brain Tumor Image Segmentation using Box-Unet Architecture

Identifying brain tumors from images which are MRI, is a very hard task and requires an experienced radiologist. The radiologist needs to recognize the tumor structure, location, and size of the brain tumor. Because of specific complications and the existence of numerous types of tumor tissues, this recognition procedure becomes complicated. Unet architecture is one of the most widely utilized architectures in segmenting brain tumor. This paper proposes a novel architecture of Unet known as Box-Unet that uses the BRATS 2020 dataset. In this architecture, boxes have been added in the upsampling part of the traditional Unet Architecture. The addition of boxes, which is a set of one or more pairs of upsampling and convolution operations, enhances the performance of segmentation of brain tumor, as compared to traditional Unet architecture in terms of sensitivity, precision, loss, accuracy, mean IoU and specificity. BOX-Unet architecture improved the Jaccard coefficient by 1.56% as compared to basic Unet Architecture.