| 摘 要: 医学超声图像常面临高噪声和边缘模糊等特点,传统的图像分割方法因采用固定的跳跃连接方法,导致出现分割图像的分辨率降低和部分细节丢失等问题。针对上述问题,本文提出一种基于双编码器和 Hi-Lo注意力机制的 Mamba-HL-UNet分割模型。首先,为了更好地提取特征图像的细节以及长距离依赖关系,主干网络部分引入MambaVSS编码器和ResNet50编码器。其次,在跳跃连接中引入 Hi-Lo注意力机制,通过调节head数量在特征图像中关注局部细节和全局信息。相比于原 UNet模型,本模型在BUSI公开数据集中获得的均交并比、像素准确度以及骰子系数分别提高了4.56%、4.12%和5.30%,由此验证了本模型的有效性和可行性。 |
| 关键词: 医学超声图像 图像分割 UNet |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 江苏省高等教育机构基础科学(自然科学)研究项目(21KJB480012) |
|
| A Medical Ultrasound Image Segmentation Method Based on Mamba-HL-UNet |
|
XUE Guoqiang1, WANG Keqing1,2, ZHANG Qiang1
|
(1. School of Automation, Nanjing University of Information Science & Technology, Nanjing 210044, China;
2. School of Automation, Wuxi University, Wuxi 214105, China)
xgq0701@163.com; wangkeqing93@sina.com; zhangqiang6426@163.com
|
| Abstract: Medical ultrasound images exhibit characteristics such as high noise and blurred edges. Traditional image segmentation methods rely on fixed skip-connection approaches, leading to reduced resolution and loss of fine details in segmented images. To address these issues, this paper proposes a Mamba-HL-UNet segmentation model based on a dua-l encoder architecture and the H-i Lo attention mechanism. Firstly, to better extract detailed features and longrange dependencies, the backbone network incorporates a Mamba VSS encoder and a ResNet50 encoder. Secondly, the H-i Lo attention mechanism is introduced in skip connections, which adjusts the number of heads to focus on both local details and global information in feature maps. Evaluated on the BUSI public dataset, the proposed model demonstrates improvements of 4.56% , 4.12% , and 5.30% in mean Intersection over Union (mIoU), pixel accuracy, and Dice coefficient compared to the original UNet model, thereby validating its effectiveness and feasibility. |
| Keywords: medical ultrasound images image segmentation UNet |
