基于3D U-Net的轻量级脑肿瘤分割网络北大核心CSCD

针对现有脑肿瘤核磁共振成像(magnetic resonance imaging, MRI)分割神经网络的参数量和计算量较大且对肿瘤区域小目标分割精度不高的问题,提出一种改进的轻量级脑肿瘤分割网络MF-RES2Net(multiple fiber residual-like networks)。该网络以3D U-Net为基础架构,将多纤模块(multi-fiber, MF)和类残差模块(RES2)相结合代替传统卷积模块。MF将特征图像的通道进行混合,增加了通道间信息的交流融合;RES2将通道均分,单通道的卷积结果相加到相邻通道,在扩大图像感受野的同时保留了细节特征,同时降低网络参数量。此外,为改善数据不平衡问题,提出一种改进的加权损失函数,提高了网络对小目标的分割精度。将MF-RES2Net在BRATS 2019数据集进行验证,完整肿瘤、核心肿瘤和增强肿瘤分割的平均Dice系数分别为89.98%、84.02%、77.62%,参数量和浮点数分别为3.16 M和16.24 G,结果表明:该网络在降低参数量和计算量的同时进一步提升了分割性能,有效地降低了网络运行时的设备要求。

Lightweight network in brain tumor segmentation based on 3D U-Net

Considering that the current neural ne tworks have some problems in brain tumor magnetic resonance imaging (MRI) segmentation,which are a large numb er of parameters and low accuracy of small target segmentation, an improved lightweight brain tumor segmentation network multiple fiber residual-like networks (MF-RES2Net) is proposed .The network is based on 3D U-Net and replaces the traditional convolution module with the multi-fiber (M F) unit and the RES2 unit.The MF unit mixes the channels of the feature image,which increases the com munication between channels.The RES2 unit divides the channels equally,and the convolution result of one single channel is added to the adjacent channels,which expands the image recept ive field and reduces the parameters while retaining feature details.In addition,a improved weighted-loss funct ion is proposed to address the network segmentation accuracy of small targets for the data imbalanc e problem. MF-RES2Net is verified on the BRATS 2019 data set,and the average Dice coeffic ients of tumor segmentation in whole tumor,core tumor and enhanced tumor region have reached 8 9.98%,84.02%, 77.62% respectively,and the network has 3.16 M parameters and 16.24 G FLOPs.The re sult shows MF-RES2Net achieves more accurate target segmentation with lower parameters and calculations, effectively reducing equipment requirements during network running.