基于双重注意力与ResNet50的REC-UNet肝实质分割模型

REC-UNet liver parenchyma segmentation model based on dual attention mechanism and ResNet50

摘要

针对肝脏超声图像中噪声干扰显著、边缘信息模糊以及空间定位难等问题，提出一种REC-UNet模型，验证其在肝脏超声图像中提取肝实质的分割精度。引入双重注意力机制与ResNet50残差网络构建REC-UNet模型，利用混合损失函数对587幅肝脏超声图像进行训练，基于消融实验对比改进模型与主流网络模型的评价指标。改进模型的平均交并比MiOU、精确率P(Precision)、召回率R(Recall)以及Dice系数分别达到了91.67%、91.58%、93.23%和91.58%，全局Dice系数较U-Net、UNet++、Attention-UNet以及FCN模型分别提高了5.10%、4.60%、4.44%以及6.13%。REC-UNet模型可以精准地分割出肝脏超声图像中的肝实质，各项分割指标相较其他主流神经网络模型均有显著提升。

Abstract

Aiming at the problems of significant noise interference, blurred edge information and difficult spatial positioning in liver ultrasound images, a REC-UNet model is proposed to verify its segmentation accuracy in extracting liver parenchyma from liver ultrasound images. The REC-UNet model was constructed by integrating the dual attention mechanism with the ResNet50 residual network. 587 liver ultrasound images were trained using a mixed loss function. Based on ablation experiments, the evaluation indicators of the improved model were compared with those of the mainstream network models. The mean intersection over union (MiOU), precision, recall rate and Dice coefficient of the improved model reached 91.67%, 91.58%, 93.23% and 91.58% respectively. The global Dice coefficient of this model was 5.10%, 4.60%, 4.44% and 6.13% higher than that of the U-Net, UNet++, Attention-UNet and FCN models, respectively. The REC-UNet model can precisely segment the liver parenchyma in ultrasound images of the liver, and its various segmentation indicators have significantly improved compared with other mainstream neural network models.

关键词

深度学习 / 肝脏超声 / 双重注意力机制 / 图像分割 / 卷积神经网络

Key words

deep learning / liver ultrasound / dual attention mechanism / image segmentation / convolutional neural network

参考文献

[1] 吕珂. 人工智能在超声检查肝脏疾病中的应用:现状、挑战及展望[J]. 中国医学影像技术, 2025, 41(8): 1327-1332.

[2] 宋丹琳, 郑静, 倪志鹏, 等. 超声图像人工智能辅助对甲状腺结节定性诊断价值的研究[J]. 中国临床医学影像杂志, 2023, 34(2): 87-90.

[3] 尹诗, 施振扬, 吴梦麟, 等. 基于深度学习的肾脏超声图像分割:现状与挑战[J]. 计算机科学, 2025, 52(9): 16-24.

[4] ZULUAGA GOMEZ J, ALMASRY Z, BENAGGOUNE K, et al. A CNN—based methodology for breast cancer diagnosis using thermal images[J]. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2021, 9(2): 131-145.

[5] 张涛, 禹宝庆, 雷鸣, 等. Dense—UNet:基于密集连接U—Net的创伤伤口图像分割[J]. 中国体视学与图像分析, 2024, 29(4): 282-289.

[6] 扈蕴琨, 王晓艳, 王秀娟. 基于级联DDR—UNet++的肝脏肿瘤图像分割方法[J]. 中国医学物理学杂志, 2025, 42(7): 901-910.

[7] 云飞, 殷雁君. 改进Attention—UNet的多尺度肝脏CT图像分割[J]. 内蒙古师范大学学报(自然科学汉文版), 2023, 52(2): 175-180.

[8] 来金玉, 钟李长, 师琳, 等. 基于二维超声图像的深度学习模型术前预测乳腺癌脉管浸润:一项人机对比研究[J]. 肿瘤影像学, 2025, 34(3): 201-207.

[9] 闫梦婷. ResNet50卷积神经网络在肺部感染图像识别中的应用与优化研究[J]. 信息与电脑, 2025, 37(23): 48-50.

[10] WANG Q, WU B, ZHU P, et al. ECA—Net: Efficient channel attention for deep convolutional neural networks[C]// Conference on Computer Vision and Pattern Recognition, Seattle: IEEE/CVF, 2020: 11534-11542.

[11] HOU Q, ZHOU D, FENG J. Coordinate attention for efficient mobile network design[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville, TN: IEEE, 2021: 13713-13722.

[12] LI X, QIN X, HUANG C, et al. SUnet: A multi—organ segmentation network based on multiple attention[J]. Computers in Biology and Medicine, 2023, 167: 107596.

[13] 于双双, 康帅, 张建军, 等. 基于改进Unet++的双时相遥感图像建筑物损伤评估模型[J]. 地球信息科学学报, 2025, 27(11): 2741-2755.

[14] 李阳. 基于VAN_ResUNet的肝脏CT图像分割[J]. 软件导刊, 2025, 24(8): 188-195.

[15] 田普光. 基于混合损失函数的U—Net网络建筑物提取[J]. 测绘与空地理信息, 2023, 46(12): 109-112.

[16] 张晓宇, 王彬, 安卫超, 等. 基于融合损失函数的3D U—Net++脑胶质瘤分割网络[J]. 计算机科学, 2021, 48(9): 187-193.

[17] MUSTHAFA N, MASUD M M, MEMON Q, et al. VISION: View—specific integrated segmentation—classification framework for accurate brain tumor detection in MRI scans[J]. PloS ONE, 2025, 20(10): e0332395.

[18] 赵珊, 郑爱玲. 判别相关分析双注意力机制的目标检测算法[J]. 计算机工程与应用, 2022, 58(17): 120-129.

[19] 高继蕊, 李华锋, 张亚飞, 等. 双注意力引导的细节和结构信息融合图像去雾网络[J]. 电子学报, 2023, 51(1): 160-11.

基金资助

山东省自然科学基金青年项目(ZR2021QH346)