基于改进U-Net和注意力机制的肝脏肿瘤图像分割

Liver Tumor Image Segmentation Based on an Improved U-Net and Attention Mechanisms

摘要: 针对肝脏肿瘤图像中病灶形态复杂、与周围组织边界模糊，导致分割精度受限的问题，提出一种基于改进U-Net和注意力机制的肝脏肿瘤图像分割模型(DMCAU-Net)。该模型通过设计双分支多尺度残差卷积模块替代传统卷积结构，利用多分支方式提取不同感受野下的特征信息，从而增强对多尺度肿瘤特征的表达能力。同时，在跳跃连接中引入CBAM双重注意力机制，对通道维和空间维特征进行自适应加权，有效抑制背景噪声干扰，使模型更加关注肝脏肿瘤关键区域。此外，构建了结合二元交叉熵损失和Dice损失的混合损失函数，以进一步平衡分类性能与分割精度。实验结果表明，与基线模型U-Net相比，所提出模型在IoU和DSC指标上分别达到90.07%和94.78%，分别提升了18.73%和11.51%。结果验证了该模型在肝脏肿瘤图像分割任务中的有效性，可为临床诊断与治疗提供可靠的技术支持。

Abstract: Aiming at the problem that the segmentation accuracy of liver tumor images is limited due to the complex morphology of lesions and the blurred boundaries between tumors and surrounding tissues, a liver tumor segmentation model based on an improved U-Net and attention mechanisms (DMCAU-Net) is proposed. In this model, a dual-branch multiscale residual convolution module is designed to replace the traditional convolution module, and feature information with different receptive fields is obtained through a multi-branch structure, thereby enhancing the extraction and representation ability of multiscale tumor features. At the same time, a CBAM dual attention mechanism is embedded into the skip connections, and the feature fusion is optimized by dynamically adjusting channel and spatial weights, which suppresses background noise interference and enables the model to accurately focus on tumor regions. In addition, a hybrid loss function combining binary cross-entropy loss and Dice loss is constructed to further balance classification performance and segmentation quality. Experimental results show that, compared with the base-line model U-Net, the IoU and DSC of the proposed model reach 90.07% and 94.78%, respectively, with improvements of 18.73% and 11.51%, respectively. The proposed model significantly improves the segmentation accuracy of liver tumors and provides reliable evidence for clinical diagnosis and treatment.