基于轻量化U-Net模型分割乳腺癌超声图像  

Lightweight U-Net model for segmentation of breast cancer ultrasound images

作　　者：张睿智 杨冲 徐栋[4] ZHANG Ruizhi;YANG Chong;XU Dong(Graduate School of Zhejiang Chinese Medical University,Hangzhou 310053,China;Department of Ultrasound,Putuo District People’s Hospital of Zhoushan City,Zhoushan 316100,China;College of Mechanical and Electrical Engineering,Hainan University,Haikou 570228,China;Department of Ultrasound,Zhejiang Cancer Hospital,Hangzhou 310022,China)

机构地区：[1]浙江中医药大学研究生院,浙江杭州310053 [2]舟山市普陀区人民医院超声科,浙江舟山316100 [3]海南大学机电工程学院,海南海口570228 [4]浙江省肿瘤医院超声医学科,浙江杭州310022

出　　处：《中国介入影像与治疗学》2024年第10期618-623,共6页Chinese Journal of Interventional Imaging and Therapy

摘　　要：目的观察轻量化U-Net(L-U-Net)模型分割乳腺癌超声图像的价值。方法回顾性分析779例女性乳腺癌患者共1009幅超声图像,以其中807幅为训练集、202幅为测试集。将MobileNetV2模块和MobileViT模块嵌入U-Net模型编码端构建L-U-Net模型,包括常规轻量L-U-Net(L-U-Net 1)和次轻量L-U-Net(L-U-Net 2)模型;以医师人工标注病灶区域结果为参考标准,评估L-U-Net模型的分割精度及轻量化程度。结果L-U-Net模型分割乳腺癌超声图像的像素准确率、交并比和戴斯相似系数与U-Net模型相近,且参数量、浮点运算数和内存使用量低于而推理时间高于U-Net模型。U-Net及L-U-Net模型分割病灶边界清晰的乳腺癌超声图像的效能均较好;对病灶边界模糊但仍可辨识的图像,U-Net模型易误标非病变区域,而L-U-Net模型分割结果更为准确;对病灶边界模糊且肉眼难以辨识的图像,3种模型均存在标注不全情况,其中U-Net和L-U-Net 1模型缺失区域更大、而L-U-Net 2模型缺失区域较小。结论利用L-U-Net 2模型分割乳腺癌超声图像能以良好的轻量化程度获得较好的分割精度。Objective To observe the value of lightweight U-Net(L-U-Net)model for segmentation of breast cancer ultrasound images.Methods A total of 1009 ultrasound images of 779 female cases with breast cancer were retrospectively analyzed,including 807 images in training set and 202 images in test set.MobileNetV2 and MobileViT modules were embedded into encoding end of U-Net model to construct L-U-Net models,i.e.conventional lightweight LU-Net(L-U-Net 1)and sub lightweight L-U-Net(L-U-Net 2)models.The segmentation accuracy and lightweight degree of L-U-Net models were evaluated taken manually annotating lesion areas by physicians as the reference standards.Results The pixel accuracy,intersection over union and Dice similarity coefficient of L-U-Net models for segmentation of breast cancer ultrasound images were similar to those of U-Net model,and the number of parameters,floating point operation and memory usage of L-U-Net model were lower but inference time were higher than those of U-Net model.UNet and L-U-Net models had better segmentation efficacy for ultrasound images of breast cancer with clear boundaries.For images with blurred lesion boundaries but still recognizable,U-Net model was prone to mislabeling non lesion areas,while L-U-Net models could provide more accurate segmentation results.For images with blurred lesion boundaries difficult to identify with naked eyes,all 3 models had incomplete segmentation,among which U-Net and L-U-Net 1 models had larger missing areas but L-U-Net 2 model had smaller missing areas.Conclusion L-U-Net 2 model could be used for segmentation of breast cancer ultrasound images with good lightweight degree and segmentation accuracy.

关 键 词：乳腺肿瘤 超声检查 神经网络 计算机 人工智能 

分 类 号：R737.9[医药卫生—肿瘤] R445.1[医药卫生—临床医学]