融合注意力和Inception模块的脑白质病变端到端分割

End-to-End Segmentation of Brain White Matter Hyperintensities Combining Attention and Inception Modules

针对目前磁共振脑影像上脑白质高信号区域的自动分割存在分割精度较低和细小病灶易漏识等问题,提出一种融合注意力和Inception的U-Net分割模型。在U-Net的编码阶段加入Inception模块以增加网络宽度,使其具有多尺度特征的提取能力,并加入注意力模块以增强网络对分割目标的关注度,两者的加入和融合可以有效提升网络的特征提取和表达能力。同时,在解码阶段的每一个卷积层上增加残差连接,可以提高网络的优化速度。此外,针对样本不均衡易导致分割结果中假阴性过多的问题,采用具有均衡调节能力的Tversky损失函数优化网络训练。实验结果表明,所提方法能够较好地分割脑白质的高信号区域,特别是小病灶区,且各项分割指标优于多个对比方法。

Aiming at solving the problems of low segmentation accuracy in the automatic segmentation of brain white matter hyperintensity region on magnetic resonance imaging brain images and easy to miss small lesions,a U-Net segmentation model combining attention and inception is proposed.In the coding stage of U-NET,the Inception module is added to increase the width of the network,so that it has the ability to extract multi-scale features,and the attention module is added to enhance the attention of the network to the segmentation target.The addition and fusion of the two can effectively improve the feature extraction and expression capabilities of the network.Simultaneously,adding residual connections on each convolutional layer in the decoding stage can improve the optimization speed of the network.In addition,because of the problem that sample imbalance easily leads to too many false negatives in the segmentation results,the Tversky loss function with balance adjustment ability is employed to optimize network training.The experimental results show that the proposed method can segment brain white matter hyperintensity region,especially the small lesion area,and each segmentation index is better than those of multiple comparison methods.