| Summary: | Tumor delineation is critical for the precise diagnosis and treatment of glioma patients. Since manual segmentation is time-consuming and tedious, automatic segmentation is desired. With the advent of convolution neural network (CNN), tremendous CNN models have been proposed for medical image segmentation. However, the small size of kernel limits the shape of the receptive view, omitting the global information. To utilize the intrinsic features of brain anatomical structure, we propose a modified U-Net with an attention block (AttU-Net) to tract the complementary information from the whole image. The proposed attention block can be easily added to any segmentation backbones, which improved the Dice score by 5%. We evaluated our approach on the dataset of BraTS 2021 challenge and achieved promising performance on this dataset. The Dice scores of enhancing tumor, tumor core, and whole tumor segmentation are 0.793, 0.819, and 0.879, respectively.
|
|---|