脑肿瘤分割算法研究

由于医学成像技术的复杂性和胶质瘤表面高度的异质性,对人脑神经胶质瘤的图像分割是医学影像分析中最具挑战性的任务之一。本研究对UNet++医学图像分割网络进行了改进,改进后的网络能够融合全尺度下的粗粒度语义和细粒度语义信息。在公开的BraTS脑肿瘤分割数据集中的335例图像上进行分割实验,采用2D与3D对比分割实验综合评价改进后网络的分割性能,将分割结果与UNet,UNet++,UNet3+医学图像分割网络的结果进行对比。在Dice相似度系数(DSC),95%Hausdorff表面距离(HSD95),Sensitivity,PPV等4个指标基础上,2D对比分割实现的指标均值分别为:83.70%,1.7,88.40%,84.96%;3D对比分割实验的指标均值为:90.79%,0.242,91.23%,91.06%。实验结果表明,改进的算法使神经胶质瘤的分割结果与金标准在区域上有更多的重叠,可以更好的完成脑胶质瘤的分割。在临床应用中,可望帮助神经外科医生高效地分离脑肿瘤与人脑周围组织,从而实现快速的计算机诊疗。

Research on Brain Glioma Segmentation Algorithm

Due to the complexity of medical imaging and the high heterogeneity of the surface of gliomas, image segmentation of human brain gliomas is one of the most challenging tasks in medical image analysis. This paper aimed to improve the UNet++ medical image segmentation network, the improved network can fuse coarse-grained semantics and fine-grained semantics at full scale. Experiments were performed on 335 images obtained from the public BraTS brain tumor segmentation data set, using 2D and 3D comparative segmentation experiments to comprehensively evaluate the segmentation performance of the improved network and compare the segmentation results with the results of UNet, UNet++, and UNet3+ medical image segmentation networks. Among the four indicators of Dice similarity coefficient (DSC), 95% Hausdorff surface distance (HSD95),sensitivity, and positive predictive value (PPV), 2D contrast segmentation achieved the mean values of the indicators of 83.70%, 1.7, 88.40%, 84.96% respectively; the mean values of the 3D contrast segmentation reached 90.79%, 0.242, 91.23%, 91.06% respectively. Compared with the segmentation result indicators of the other three networks, in the 2D comparison experiment, DSC increased by 1.82% on average, HSD95 decreased by 0.35 on average, sensitivity increased by 2.13% on average, and PPV increased by 0.80% on average; in the 3D comparison experiment, DSC increased by 2.78% on average, HSD95 decreased by 0.076 on average, Sensitivity increased by 3.81% on average, and PPV increased by 0.68% on average. It was shown that the proposed algorithm made the segmentation result of glioma and the gold standard overlap more in the region, and completed the segmentation of glioma better. It is expected to help neurosurgeons to more precisely separate brain tumors and tissues around the brain and achieve rapid computer diagnosis and treatment.