Abstract: 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.

Key words: 3D U-Net++, Fusion loss function, Glioma, Multimodal magnetic resonance imaging, Tumor segmentation