基于改进U-Net的肝脏肿瘤图像分割算法

Liver Tumor Image Segmentation Algorithm Based on Improved U-Net

摘要: 肝脏肿瘤是肝脏疾病领域发病率高、致死率突出的恶性病变，因此肝脏肿瘤的早期筛查工作对于降低肿瘤恶化概率和病患的死亡率有着至关重要的意义。基于深度学习的肝脏肿瘤图像计算机辅助诊断技术，对优化临床诊疗流程、改善患者预后具有重要意义。针对传统医学图像分割网络在肝脏图像分割任务中深层特征提取能力不足、有效特征关注度低，进而导致分割精度欠佳的问题，本文提出一种改进型U-Net网络模型EFF-UNet。该模型在U-Net基础架构上，在跳跃连接处嵌入了EFF模块，该模块整合了高效通道注意力(ECA)、空间注意力(SA)、高效注意力门控(EAG)三种注意力机制，核心目标是解决传统U-Net跳连过程中浅层/深层特征融合不充分、有效特征关注度低的问题。在肝脏图像分割专用数据集上的实验结果表明，EFF-UNet模型的Dice相似系数与平均交并比(mIoU)分别达到72.36%和69.32%，相较于原始U-Net模型，两项核心分割指标分别提升2.42个百分点和1.74个百分点。研究结果证实，EFF-UNet可有效提升肝脏肿瘤图像分割精度，为临床肝脏肿瘤辅助诊断提供了一种更具潜力的技术方案。

Abstract: Liver tumors are malignant lesions with high incidence and prominent mortality in the field of liver diseases. Therefore, early screening of liver tumors is of crucial significance for reducing the probability of tumor progression and the mortality rate of patients. Deep learning-based computer-aided diagnosis technology for liver tumor images plays an important role in optimizing clinical diagnosis and treatment processes and improving patient prognosis. Aiming at the problems that traditional medical image segmentation networks have insufficient deep feature extraction capability and low attention to effective features in liver image segmentation tasks, which further lead to poor segmentation accuracy, this paper proposes an improved U-Net network model named EFF-UNet. On the basis of the U-Net basic architecture, this model embeds an EFF module at the skip connections. The module integrates three attention mechanisms, namely Efficient Channel Attention (ECA), Spatial Attention (SA) and Efficient Attention Gating (EAG). Its core goal is to solve the problems of insufficient fusion of shallow and deep features and low attention to effective features during the skip connection process of the traditional U-Net. Experimental results on a dedicated dataset for liver image segmentation show that the Dice Similarity Coefficient and mean Intersection over Union (mIoU) of the EFF-UNet model reach 72.36% and 69.32% respectively. Compared with the original U-Net model, the two core segmentation metrics are improved by 2.42 percentage points and 1.74 percentage points respectively. The research results confirm that EFF-UNet can effectively improve the segmentation accuracy of liver tumor images, providing a more potential technical solution for clinical computer-aided diagnosis of liver tumors.