Abstract:Precise diagnosis of brain tumors has important significance for improving the survival rate of patients and providing
positive and effective treatment. Magnetic resonance (MR) imaging can provide the diagnosis of brain tumors and increase the
rate of diagnosis of brain tumor. Accurate segmentation of 3D brain tumor MR images is of great significance for the diagnosis,
treatment and postoperative tracking of brain tumors. In this paper, based on the brain tumor segmentation on 3D MR images,
a deep neural network algorithm for directly optimizing the evaluation index’s (Sorensen Dice coefficient) new loss function
was proposed, which could directly optimize the three important clinical requirements: whole tumor area, tumor core area and
enhanced tumor area. The average Sorensen Dice coefficients of the final test set in the three target regions of the whole tumor
area, the tumor core area, and the enhanced tumor area were 0.875, 0.829, and 0.695, respectively, which were better than the
traditional cross entropy loss function. It provides new automa ted tools for the accurate segmentation of brain tumors.
刘昊,王冠华,章强,李雨泽,陈慧军. 3D脑肿瘤分割的Dice损失函数的优化[J]. 中国医疗设备, 2019, 34(5): 20-23.
LIU Hao, WANG Guanhua, ZHANG Qiang, LI Yuze, CHEN Huijun. Optimization of Dice Loss Function for 3D Brain Tumor Segmentation. China Medical Devices, 2019, 34(5): 20-23.