用于颈部超声图像的SED-UNet分割方法研究
Research on SED-UNet Segmentation Method for Neck Ultrasound Image
刘明珠 1付聪 1宋诗杰 1赵首博1
作者信息
- 1. 哈尔滨理工大学 测控技术与仪器黑龙江省高校重点实验室,哈尔滨 150080
- 折叠
摘要
超声图像作为目前常用的医疗诊断手段之一,人工判读超声图像很大程度上依赖于医生主观经验知识,耗时耗力,难以满足快速、批量的临床诊断需求,因此提出了一种基于深度学习和可变形卷积U-Net的图像分割模型SED-UNet.用可变形卷积结合BN和Dropout层对原网络的卷积运算进行优化改进,提升网络收敛性、增加网络模型的鲁棒性、提升模型的训练效率,用SENet模块在解码阶段的跳跃连接处进行优化改进,提升分割准确率,进而构建适用于颈部超声图像分割的卷积神经网络模型.测试结果表明,提出的 SED-UNet 模型在颈部超声图像的自动分割方面性能良好,F1 系数、精确率、MIoU参数相比传统U-Net结构分别提升了3.94%、7.61%、7.15%,从客观评价指标上达到了较好的分割效果.
Abstract
Ultrasound image is one of the commonly used medical diagnosis methods.Manual interpretation of ultrasound image largely depends on doctors′ subjective experience and knowledge,which is time-consuming and labor-consuming,and is difficult to meet the needs of rapid and batch clinical diagnosis.Therefore,this paper proposes an image segmentation model SED-UNet based on deep learning and deformable convolution U-Net.The deformable convolution combined with BN and Dropout layer is used to optimize and improve the convolution operation of the original network,improve the network convergence,increase the robustness of the network model and improve the training efficiency of the model.The senet module is used to optimize and improve the jump connection in the decoding stage,improve the segmentation accuracy,and then construct a convolution neural network model suitable for neck ultrasound image segmentation.The test results show that the SED-UNet model proposed in this paper has good performance in the automatic segmentationofneck ultrasound images.The F1 coefficient,accuracy and MIoU parameters are improved by 3.94%,7.61%and 7.15%respectively compared with the traditional U-Net,and achieve a better segmentation effect from the objective evaluation index.
关键词
SENet模块/U-Net/可变形卷积/图像分割Key words
SENet module/U-Net/deformable convolution/image segmentation引用本文复制引用
基金项目
国家自然科学基金青年基金(61801148)
出版年
2024
NETL
NSTL
万方数据