Abstract

Info

Title:: Fully automatic segmentation algorithm of brain tumor based on three-dimensional UNet and hybrid focal loss function

Author(s):: TIAN Hengyi; XIAO Hongbing; JI Yarong; Rahman Md Mostafizur; School of Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China

Keywords:: Keywords: brain tumor deep learning 3D UNet hybrid focus loss function

Abstract:: Abstract: A fully automatic brain tumor segmentation algorithm combining hybrid focus loss function and 3D UNet is proposed to address the problem of low segmentation accuracy caused by the serious data imbalance of different categories such as normal brain tissue and brain tumor. In 3D UNet framework, the hybrid focus loss function including focus loss and improved focus Tversky loss which possess complementary advantages can effectively alleviate the adverse effects caused by the class imbalance of input and output data, and make the segmentation model focus on the samples which are difficult to be classified and learned. The experiments on public brain tumor dataset show that the proposed segmentation model with hybrid focal loss function can achieve a mean Dice coefficient of 89.01%, 88.67%, and 83.74% in whole tumor, tumor core and enhanced tumor, respectively, and that the mean Hausdorff distances were 14.29, 5.01 and 3.84 mm, respectively. The segmentation model based on hybrid focal loss function can significantly improve the segmentation accuracy of tumor core and enhanced tumor which are difficult to be classified due to class imbalance.