Abstract: Automatic segmentation of the liver from the patient’s abdominal CT images is significant for the diagnosis of liver disease. Since the bottom-up feature fusion methods in U-Net ignore the importance of low-level features, the segmentation results are not satisfactory. In addition, the gray values for the liver and its adjacent organs are very similar, it is therefore difficult to distinguish some tiny detail features. To address the above problems, a liver segmentation network based on multi-scale semantic features fusion and attention mechanism (MSFA-Net) is proposed. Firstly, dilated residual convolution (DRC) is used to capture multi-scale features. Then, the MSFA module combines top-down and bottom-up multi-scale feature fusion methods with the attention mecha-nism to fully fuse multi-scale features and pay attention to tiny features. Finally, the feature maps are summed by deep supervise (DS) to improve the segmentation effect. The ablation study is performed on the MICCAI2017 LiTS and 3DIRCADb datasets. The dice per case (DC) and dice global (DG) scores of 0.961 and 0.965 are obtained on the LiTS dataset, with an improvement of 3.4% and 2.0% respectively compared with baseline network. Both DC and DG scores of 0.965 are obtained on the 3DIRCADb dataset, with an increase of 3.5% and 3.3% respectively compared with baseline network.

Key words: liver segmentation, U-Net, multi-scale feature fusion, attention mechanism, deep supervision

摘要： 从患者的腹部CT图像中自动分割出肝脏对于肝脏疾病的诊断意义重大。由于在U-Net中使用自下而上的特征融合方式忽略了低级特征的重要性，导致网络分割性能较差，以及肝脏与相邻器官组织的灰度值较为相似，使得一些微小的细节特征不易被关注。针对以上问题，提出了一种基于多尺度语义特征融合和注意力机制的肝脏分割网络（MSFA-Net）。首先，使用空洞残差卷积（DRC）捕获多尺度特征；然后，采用MSFA模块将自上而下和自下而上的多尺度特征融合方法与注意力机制相结合，来充分融合多尺度特征和关注微小特征；最后，通过深度监督（DS）对特征图求和来提升分割效果。在MICCAI2017 LiTS和3DIRCADb数据集上进行了消融研究，在LiTS数据集上获得了0.961和0.965的DC和DG评分，比基线网络分别提高了3.4%和2.0%；在3DIRCADb数据集上，其DC和DG评分同为0.965，比基线网络分别提高了3.5%和3.3%。

关键词: 肝脏分割, U-Net, 多尺度特征融合, 注意力机制, 深度监督