DFU-Net:基于多尺度特征融合的肝脏肿瘤分割网络  

DFU-Net: A Liver Tumor Segmentation Network Based on Multi-Scale Feature Fusion

作　　者：佘海州 高凝 李鹤[2] 

机构地区：[1]沈阳航空航天大学电子信息工程学院,辽宁 沈阳 [2]沈阳工学院信息与控制学院,辽宁 抚顺

出　　处：《图像与信号处理》2025年第1期62-73,共12页Journal of Image and Signal Processing

摘　　要：肝脏以及肝脏肿瘤的有效分割是肝部疾病在临床诊断的关键步骤。文章针对肝脏结构复杂、肝脏与相邻器官像素强度差异小、肝脏边界模糊等特点,提出了一种可以进行多尺度特征融合的肝脏肿瘤分割网络。该方法根据肝脏CT图像特点,在3D U-Net的基础上进行改进,提升了网络提取特征的感受野,减少了传递过程中信息的丢失。同时,在网络中引入密集融合模块,该模块可对不同尺度下的特征图进行特征融合,通过边缘信息和差异信息的融合来提升网络信息提取的性能,避免传递过程中肿瘤部分等小目标特征的丢失。在LiTS17数据集上的实验结果表明,该模型对肝脏分割的Dice系数达到了0.9504,对肿瘤分割的Dice系数达到了0.7046,实验结果证明了该方法的出色分割性能和有效性。Effective segmentation of the liver and liver tumors is a key step in the clinical diagnosis of liver diseases. This paper addresses the complexity of liver structure, the small difference in pixel intensity between the liver and adjacent organs, and the vagueness of liver boundaries, proposing a liver tumor segmentation network capable of multi-scale feature fusion. Based on the characteristics of liver CT images, this method improves upon the 3D U-Net, enhancing the network’s receptive field for feature extraction and reducing information loss during transmission. At the same time, a dense fusion module is introduced into the network, which can fuse feature maps at different scales, enhancing the network’s performance in information extraction through the integration of edge and difference information and preventing the loss of small target features such as tumor parts during transmission. Experimental results on the LiTS17 dataset show that the model achieved a Dice coefficient of 0.9504 for liver segmentation and 0.7046 for tumor segmentation, demonstrating the excellent segmentation performance and effectiveness of this method.

关 键 词：CT图像分割 3D U-Net 多尺度特征融合 密集连接 

分 类 号：R73[医药卫生—肿瘤]