Brain Tumor Segmentation Using Zernike Moments in U-Net

Abstract

The paper proposes fully automated brain tumor segmentation using Zernike moments as an initial feature in U-Net instead of a random kernel. Recent studies have shown Convolutional neural networks gained momentum in image segmentation due to an increase in computation power and availability of a large number of datasets. Among Convolutional Neural Networks, U-Net is most extensively used in medical image segmentation due to its high-resolution retaining capability. In this document, MRI images of brain tumors are segmented by varying the moment’s order in Zernike moments as initial kernels to the U-Net. Zernike moments are used to extract shape information from the brain MRI, its multi-level configuration is useful for hierarchical feature learning in U-Net. Th is model yielded a Dice score of 0.85, 0.88, and 0.81 for core, whole tumor, and enhancing tumor respectively.