一种基于改进U-Net的肝脏肿瘤分割方法

A Liver Tumor Segmentation Method Based on Improved U-Net

肝脏肿瘤分割是肝癌诊断与治疗不可或缺的重要环节。针对传统的U-Net网络在形状、大小、位置复杂多变且边界模糊的肿瘤分割中信息丢失、分割精度低等问题,对其进行改进以提高肝脏肿瘤分割精度。首先,在编码阶段使用混合空洞卷积替换原有卷积块,增大感受野、获取更多的上下文信息;在解码阶段采用密集上采样卷积,捕获和解码更详细的信息;引入残差模块,加速模型的训练并防止网络退化。其次,在每个跳跃连接之间加入注意力机制,使模型重点关注感兴趣区域,抑制冗余特征;使用组归一化(GN)代替常用的批量归一化(BN),减小Batch Size过小对网络准确性的影响,并结合Focal Tversky损失函数以改善类不平衡问题。通过LiTS2017数据集的实验表明,相较于传统U-Net,所提改进模型在肝脏和肿瘤分割中的Dice指标分别提升了3.56%和4.21%,召回率提升了3.71%和5.35%。

Segmentation of liver tumor is an indispensable link in diagnosis and treatment of hepatocellular carcinoma.Aiming at the problems of information loss and low segmentation accuracy of traditional U-Net network in tumor segmentation with complex shape,size and location and fuzzy boundary,it is improved to increase the accuracy of liver tumor segmentation.Firstly,in the encoding stage,the hybrid dilated convolution is used to replace the original convolution block to increase the receptive field and obtain more context information.In the decoding stage,dense upsampling convolution is used to capture and decode more detailed information.The residual module is introduced to speed up the training of the model and prevent network degradation.Secondly,the attention mechanism is added between each jump connection to make the model focus on the region of interest and suppress redundant features.Group Normalization(GN)is used instead of common Batch Normalization(BN)to reduce the impact of too small Batch Size on network accuracy,and combined with Focal Tversky loss function to improve the class imbalance problem.Experiments on the LiTS2017 data set show that compared with the traditional U-Net,the Dice index of the proposed improved model in liver and tumor segmentation has increased by 3.56%and 4.21%respectively,while the recall rate has increased by 3.71%and 5.35%.