Modified 3D U-Net For Brain Tumor Segmentation

Abstract

Since the brain serves as the central command and control center for the human body, brain cancer is one of the most lethal tumors. The automatic segmentation of brain tumors from multimodal images is critical in diagnosis and therapy. The difficulty in identifying the location and position of the cancer due to the different image intensity ranges is one of the challenges in segmenting brain tumors. On the other hand, the previous research architecture U-Net has limitations in terms of gathering multiscale information and delineating complicated tumor borders precisely is hard to obtain. As a result, in this study, we use Modified U-Net with add several key enhancements to improve segmentation accuracy and capture fine-grained details. In this study, we use BRATS2020, and add more augmentations to improve our model. As a consequence, our model achieved 79whole tumor, 80% tumor core, and 82% in enhanced tumor using dice score