基于融合损失函数的3D U-Net++脑胶质瘤分割网络

Glioma Segmentation Network Based on 3D U-Net++with Fusion Loss Function

胶质瘤是大脑和脊髓胶质细胞癌变所产生的、最常见的原发性颅脑肿瘤。从多模态MRI中对胶质瘤组织进行可靠的分割具有很重要的临床价值,但是由于脑胶质瘤本身及周边组织较为复杂以及浸润性导致的边界模糊等,导致对脑胶质瘤的自动分割有一定的难度。文中构建了使用融合损失函数的3D U-Net++网络来对脑胶质瘤的不同区域进行分割,该网络使用不同层级的U-Net模型进行密集嵌套连接,使用网络的4个分支的输出结果作为深度监督以更好地结合深层和浅层的特征进行分割,并结合了Dice损失函数和交叉熵损失函数作为融合损失函数来提升小区域的分割精度。在2019年多模态脑肿瘤分割挑战赛(BraTs)的公共数据集划分的独立测试集中,采用Dice系数、95%Hausdorff距离、平均交并比(mIoU)、查准率(PPV)指标对所提方法进行了评估。结果表明,全肿瘤区域、肿瘤核心区域和增强肿瘤区域的Dice系数分别为0.873,0.814,0.709;其95%Hausdorff距离分别为15.455,12.475,12.309;其mIoU分别为0.789,0.720,0.601;其PPV分别为0.898,0.846,0.735。与基础的3D U-Net以及带深度监督的3D U-Net相比,所提方法可以有效地利用多模态的深层和浅层的信息,有效利用了空间信息,同时使用了Dice系数和交叉熵的融合损失函数,从而有效提升了对肿瘤各区域的分割精度,尤其是对小面积的增强肿瘤区域的分割精度。

Glioma is the most common primary brain tumor caused by cancerous glial cells in the brain and spinal cord.Reliable segmentation of glioma tissue from multi-mode MRI is of great clinical value.However,due to the complexity of glioma itself and surrounding tissues and the blurring of boundary caused by invasion,automatic segmentation of glioma is difficult.In this paper,a 3D U-Net++network using the fusion loss function is constructed to segment different areas of glioma.The network uses different levels of U-Net models for densely nested connections,and uses the output results of the four branches of the network as depth supervision so that the combination of deep and shallow features can be better used for segmentation,and combines Dice loss function and cross entropy loss function as a fusion loss function to improve the segmentation accuracy of small regions.In the independent test set divided by the public data set of the 2019 Multimodal Brain Tumor Segmentation Challenge(BraTs),the proposed method is evaluated with Dice coefficient,95%Hausdorff distance,mIoU(mean intersection over union),and PPV(precision)indicators.The whole tumor,the core region and the enhancing tumor region of Dice coefficient are 0.873,0.814,0.709;the 95%Hausdorff coefficient are 15.455,12.475,12.309 respectively;the mIoU are 0.789,0.720,0.601 respectively;the PPV are 0.898,0.846 and 0.735 respectively.Compared with the basis of 3D U-Net and 3D U-Net with depth of supervision,the proposed method can make use of more effective modal of the deep and shallow information,effectively use the space information.And the fusion loss function combined by the dice coefficient and the cross-entropy loss function can effectively enhance tumor segmentation accuracy of each area,especially the segmentation accuracy of small tumor areas such as enhancing tumor.