基于深度迁移学习的肝肿瘤增强CT分割方法研究  

Liver tumor segmentation on enhanced CT images based on deep transfer learning

作　　者：姚瑶[1] 周艳艳 曹云堡 叶晔 陈云志[1] 罗中华 YAO Yao;ZHOU Yanyan;CAO Yunbao;YE Ye;CHEN Yunzhi;LUO Zhonghua(Hangzhou Vocational&Technical College,Hangzhou 310018,China;Interventional Center,the Second Affiliated Hospital of the Air Force Military Medical University,Xi’an 710038,China)

机构地区：[1]杭州职业技术学院,浙江杭州310018 [2]空军军医大学第二附属医院介入中心,陕西西安710038

出　　处：《实用放射学杂志》2023年第10期1712-1716,共5页Journal of Practical Radiology

基　　金：空军军医大学第二附属医院前沿交叉研究项目(2021TYJC-004);杭州市科技发展计划项目(202204B05);2021年度浙江省基础公益研究计划项目(LGG21E050005)。

摘　　要：目的探讨深度迁移学习方法辅助腹部增强CT影像在肝肿瘤分割和定量诊断中的应用价值。方法收集86例经临床确诊为肝肿瘤患者的CT影像资料,包括动脉期、门静脉期和延迟期(平衡期)增强CT影像数据。并由2位高年资放射诊断医师勾画肿瘤病灶感兴趣区(ROI),将病灶ROI平均值作为金标准。采用基于深度迁移学习的3D U-net网络模型,将门静脉期肝肿瘤CT分割任务中网络提取的语义信息迁移到动脉期及延迟期肝肿瘤CT影像分割任务中,从而避免提取类似语义信息过程的重复训练,有效提高动脉期和延迟期肝肿瘤CT影像分割性能。结果与直接采用3D U-net网络模型训练和测试相比,采用深度迁移学习的3D U-net网络模型后,动脉期及延迟期肿瘤CT影像分割结果有显著提高。结论基于深度迁移学习的3D U-net网络增强CT肝肿瘤分割和诊断模型有效提升多期肝肿瘤增强CT影像分割结果,可以辅助医生诊断和治疗。Objective To explore the application value of the deep transfer learning method in the segmentation and quantitative diagnosis of liver tumors with enhanced abdominal CT images.Methods CT images of 86 patients with liver tumors diagnosed by clinicians were collected,including arterial,portal and delayed phases.The region of interest(ROI)of the tumor was delineated by two senior radiologists,and the mean value of the lesion was taken as a a gold standard.The 3D U-net network model based on deep transfer learning was adopted to transfer semantic information extracted from the liver tumors in the portal phase to the segmentation network in the arterial and delayed phases so as to avoid repetitive training in the process of extracting similar semantic information and to improve liver tumors segmentation performance in the arterial and delayed phases effectively.Results Compared with the 3D U-net network model,the segmentation results of arterial and delayed phases tumor on CT images were significantly improved when using the 3D U-net network model with deep transfer learning.Conclusion The 3D U-net network model based on deep transfer learning for liver tumor segmentation and diagnosis can improve segmentation results of multi-phase enhanced CT images effectively and assist the diagnosis and treatment.

关 键 词：深度迁移学习 肝脏肿瘤 分割 计算机体层成像 

分 类 号：R735.7[医药卫生—肿瘤] R814.42[医药卫生—临床医学]