基于计算机体层成像影像本地化训练策略的乳腺癌临床靶区自动勾画效果研究  

Study on the Effect of Automatic Mapping of Breast Cancer Clinical Target Area Based on CT Image Localization Training Strategy

作　　者：赵德鹤 邱玉婷 董为 唐锦浩 胡应奎 陈锦树 马爽 ZHAO De-he;QIU Yu-ting;DONG Wei;TANG Jin-hao;HU Ying-kui;CHEN Jin-shu;MA Shuang(Radiotherapy Department,Shenshan Center Hospital of Sun Yat-sen Memorial Hospital,Sun Yat-sen University,Guangdong Shanwei 516600;Radiotherapy Department,Sun Yat-sen Memorial Hospital,Sun Yat-sen University,Guangdong Guangzhou 510120)

机构地区：[1]中山大学孙逸仙纪念医院深汕中心医院肿瘤科放疗专科,广东汕尾516600 [2]中山大学孙逸仙纪念医院放疗科,广东广州510120

出　　处：《中国医疗器械信息》2025年第3期47-49,135,共4页China Medical Device Information

基　　金：汕尾市科技计划(项目名称:基于深度学习的CT影像本地化训练策略对乳腺癌临床靶区自动勾画效果的影响研究,项目编号:2023C010)。

摘　　要：目的:探讨基于计算机体层成像(CT)影像本地化训练策略在乳腺癌临床靶区自动勾画系统中的应用效果。方法:共收集了150例左侧乳腺癌患者的CT图像数据,选用了U-net网络结构作为试验模型,并采用残差U-net网络模型作为对照组进行试验,对照模型命名为U-net-R。数据集1用于建立预训练模型,数据集2被收集用于测试和本地化训练。使用U-net和对比模型U-net-R进行试验,同时也对比了从头训练与本地化训练的分割结果。结果:U-net和U-net-R两种网络模型经数据集1的训练后,U-net-R模型和U-net模型均表现出性能下降的趋势,且U-net-R模型下降幅度低于U-net模型。本地化训练后的两个模型在Dice相似系数上均有显著提升,分别提高了13.30%和9.56%。自动勾画的分割结果显著优于常规自动勾画系统的勾画结果。结论:CT影像本地化训练策略有利于提高其在数据集迁移时的特征学习能力,获得更接近使用者需求的勾画结果,证明在适应新数据集方面的有效性。Objective:This study investigated the effect of CT image-based localization training strategy(localization training strategy,LTS)in the automatic delineation system of the clinical target area of breast cancer.Methods:Collecting CT image data from 150 left breast cancer patients,we selected the U-net network structure as the experimental model,and used the residual U-net network model as the control group,named U-net-R.Dataset 1 was used to build the pretraining model,and dataset 2 was collected for testing and localization training.Experiments using U-net and contrast model U-net-R,while also comparing the segmentation results from ab initio training with localization training.Results:The experimental results show that after training of dataset 1,both U-net-R and U-net-R network models show a performance decline trend,and U-net-R model has a lower performance decline than U-net model.After localization training,the Dice similarity coefficient of the two models improved significantly,by 13.30%and 9.56%respectively.The segmentation result of automatic sketching is significantly better than that of conventional automatic sketching system.Conclusion:The CT image localization training strategy is beneficial to improve its feature learning ability during dataset migration,obtain delineation results closer to user needs,and prove its effectiveness in adapting to new datasets.

关 键 词：乳腺癌 临床靶区 自动勾画 本地化训练策略 医学影像分割 

分 类 号：R445[医药卫生—影像医学与核医学]