基于双解码路径DD-UNet的脑肿瘤图像分割算法

针对医学图像中病灶区域尺度不一、边界模糊和周围组织强度不均匀所导致的分割精度降低问题，提出了一种基于双解码器的脑肿瘤图像分割模型。为了增强特征的表征力，提出了高阶微分残差模块并使用不同空洞率的扩张卷积用于提取特征编码，提高了网络模型的分割性能；引入上下文语义信息感知模块(multi scale dilation, MSD),从不同的目标尺度中提取更多的精细信息，提高了对结构细节信息的捕获能力，同时减少了编解码器之间的特征差异；在空间解码路径中使用选择性聚合空间注意力模块(spatial aggregation attention module, SAAM),增加了对有效空间特征的权重比例，减少了无效的特征干扰。在脑肿瘤数据集上进行了实验验证，实验结果表明，所提算法的Dice系数、平均交并比、敏感性、特异性、准确率等指标分别为：93.35%、90.71%、91.15%、99.94%、96.75%。

Brain tumor image segmentation algorithm based on dual decoding path DD-UNet

In view of the reduction of segmentation accuracy caused by different scales of focus area,fuzzy boundary and uneven intensity of surrounding tissues in medical images, a brain tumor image segmentation model based on double decoder is proposed.In order to enhance the representation of features,a high-order differential residual module is proposed,and the expanded convolution with different void rates is used to extract the feature coding,which improves the segmentation performance of the network model;The context semantic information perception module is introduced to extract more fine information from different target scales,which improves the ability to capture structural details and reduces the feature differences between codecs.The selective spatial aggregation attention module (SAAM) is used in the spatial decoding path to increase the weight proportion of effective spatial features and reduce the interference of invalid features. Experimental verification on different brain tumor data sets is completed. The experimental results show that the Dice coefficient,average intersection union ratio,sensitivity,specificity and accuracy of the proposed algorithm are 93.35%,90.71%,91.15%,99.94% and 96.75%,respectively.