摘要
目的 :为实现对超声图像乳腺肿瘤准确、高效地分割,提出一种融合EfficientNet和U-Net的分割方法。方法:首先将U-Net的编码器替换成EfficientNet B4中的特征提取网络,然后引入Dice损失函数和边界损失函数,再以一定权重与交叉熵损失函数组合后得到复合损失函数,最后将EfficientNet B4在数据集ImageNet上的训练权重作为预训练权重,在公开数据集Dataset B上将该方法与U-Net、Res-UNet、VGG-UNet、Dense-UNet方法对超声图像乳腺肿瘤的分割效果进行比较。结果:该方法相较其他方法分割结果更优且参数量更小,Dice相似性系数为87.46%,相较原始的U-Net(Dice相似性系数为69.75%)提升了约18%。结论:提出的方法对超声图像乳腺肿瘤具有较好的分割效果,可为乳腺超声计算机辅助诊断系统的发展奠定基础。
Objective To propose a segmentation method incorporating EfficientNet and U-Net to segment the tumor in breast ultrasound images efficiently. Methods Firstly the encoder of U-Net was replaced with the feature extraction network in EfficientNet B4, then the Dice loss function and boundary loss function were introduced, and the compound loss function was obtained after combining with the cross-entropy loss function with certain weights, finally the training weights of Efficient Net B4 on the dataset ImageNet were used as the pre-training weights. The method proposed was compared with U-Net, ResUNet, VGG-UNet and Dense-UNet methods on the publicly available Dataset B for tumor segmentation in breast ultrasound images. Results The method proposed gave better results and had lower parameter than other segmentation methods, with a Dice similarity coefficient of 87.46%, which was about 18% higher than that(69.75%) of the original U-Net. Conclusion The method proposed behaves well in tumor segmentation of breast ultrasound images, which lays a foundation for the development of the computer aided diagnosis(CAD) system for breast ultrasound.
作者
肖丹
李文彬
张红梅
XIAO Dan;LI Wen-bin;ZHANG Hong-mei(School of Life Science and Technology,Xi'an Jiaotong University,Xi'an 710049,China;School of Medical Technology,Xi'an Medical University,Xi'an 710021,China)
出处
《医疗卫生装备》
CAS
2022年第11期8-13,共6页
Chinese Medical Equipment Journal
基金
国家自然科学基金面上项目(62171366)。
