改进Unet++在脑肿瘤图像分割的研究

Research on segmentation of brain tumor image based on improved Unet++

针对计算机辅助脑肿瘤图像分割精度不高,提出改进的密集连接网络Unet++脑肿瘤自动分割网络。分别将残差块和数据相关型上采样Dupsampling融入网络的编码、解码部分,提高特征提取的能力并防止梯度消失;使用Mish激活函数代替Relu激活函数,更平滑的曲线有助于提升网络的非线性特征提取能力和泛化性;使用交叉熵和Dice结合的损失函数,进一步提升分割精度。该方法在BraTs2019部分数据上验证,在全肿瘤、核心肿瘤和增强肿瘤分割结果的Dice系数分别达到0.9236、0.8745、0.8404,豪斯多夫距离为1.806、2.994、1.865,优于大多数脑肿瘤分割模型。

Aiming at the low accuracy of computer-aided brain tumor image segmentation,an improved dense connection network Unet++brain tumor automatic segmentation network was proposed.The residual block and data correlation upsampling structure were integrated into the coding part and decoding part of the network respectively,which effectively improved the ability of feature extraction and prevented gradient from disappearing.Using the Mish activation function instead of the Relu activation function,the smoother curve improved the ability of nonlinear feature extraction and generalization of the network.A loss function combining cross-entropy and Dice was used to further improve the segmentation accuracy.The method is validated on BraTs2019 partial data,and the dice coefficients of the whole tumor,core tumor and enhanced tumor are 0.9236,0.8745 and 0.8404 respectively,and the Hausdorff distance is 1.806,2.994 and 1.865,which are better than most brain tumor segmentation models.