基于三重注意力的脑肿瘤图像分割网络

Triple Attention Segmentation Network for Brain Tumor Images

脑肿瘤图像分割问题是脑肿瘤临床诊断和治疗脑肿瘤疾病计算机辅助诊断的基础。针对脑肿瘤MRI图像分割网络深度过深和局部与全局特征信息联系匮乏导致图像分割精度降低等问题,提出一种基于三重注意力的脑肿瘤图像分割网络。首先,借鉴残差结构,将原始图像分割网络结构的编码层和解码层中的卷积模块替换为深度残差模块,解决网络加深带来的梯度消失问题。其次,通过引入三重注意力模块,融合图像局部与全局特征信息,使网络更好地学习重要的图像特征信息,提升网络对脑肿瘤图像的分割精度。最后,在MICCAI比赛发布的BraTS脑肿瘤图像分割数据集上(包括335例患者病例),采用Dice系数等脑肿瘤评价指标进行性能评估。其中,脑肿瘤整体可达85.20%,脑肿瘤核心可达87.10%,增强脑肿瘤区域可达80.80%。实验结果显示,所提出的分割网络能够在不增加计算时间的前提下提高脑肿瘤MRI图像的分割性能。

Brain tumor segmentation is the basis of clinical routine and treatment of brain tumor diseases with computer-aided diagnosis.In this paper,we proposed a triple attention segmentation network based on brain tumor images,aiming to solve the problems of the current brain tumor MRI image segmentation network that has too many layers and is lack of connection between local and global feature information,which leads to the reduction of image segmentation accuracy.First,inspired by the residual network,we replaced the convolution module both in the encoding and decoding layer of the original segmentation network with a deep residual module to solve the problem of gradient disappearance caused by network deepening.Next,by introducing a triple attention module to combine local and global image features,the network was able to learn important image features better and improved the network′s segmentation accuracy of brain tumor images.Finally,The improved network was evaluated by the Dice coefficient,and other brain tumor indicators were adopted on the BraTS brain tumor MRI image datasets released by the MICCAI competition includes 335 patient cases,among which the whole brain tumor score reached 85.20%,the brain tumor core score reached 87.10%,and the enhanced brain tumor area score reached 80.80%.Experimental results showed that the proposed segmentation network increased the segmentation performance of brain tumor MRI images without increasing the training time.