S3D-UNet: Separable 3D U-Net for Brain Tumor Segmentation

Abstract

Brain tumor is one of the leading causes of cancer death. Accurate segmentation and quantitative analysis of brain tumor are critical for diagnosis and treatment planning. Since manual segmentation is time-consuming, tedious and error-prone, a fully automatic method for brain tumor segmentation is needed. Recently, state-of-the-art approaches for brain tumor segmentation are built on fully convolutional neural networks (FCNs) using either 2D or 3D convolutions. However, 2D convolutions cannot make full use of the spatial information of volumetric medical image data, while 3D convolutions suffer from high expensive computational cost and memory demand. To address these problems, we propose a novel Separable 3D U-Net architecture using separable 3D convolutions. Preliminary results on BraTS 2018 validation set show that our proposed method achieved a mean enhancing tumor, whole tumor, and tumor core Dice scores of 0.74932, 0.89353 and 0.83093 respectively. Finally, during the testing stage we achieved competitive results with Dice scores of 0.68946, 0.83893, and 0.78347 for enhancing tumor, whole tumor, and tumor core, respectively.