北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2023, Vol. 46 ›› Issue (6): 27-0.

• 智慧医疗 • 上一篇    下一篇

基于联邦风格迁移的图像分割

马静超,印象,胡陈菲,马博渊,班晓娟   

  1. 1. 北京科技大学

    2. 河钢集团有限公司

    3. 辽宁材料实验室材料智能技术研究所

  • 收稿日期:2022-10-24 修回日期:2023-01-14 出版日期:2023-12-28 发布日期:2023-12-29
  • 通讯作者: 马博渊 E-mail:mbytony@ ustb. edu. cn。
  • 基金资助:
    国家自然科学基金项目(62106019);中央高校基本科研业务费专项项目(06500221,FRF-IDRY-21-022);佛山市科技创新专项 资金项目(BK21BF002,BK22BF010);北京科技大学青年科技创新基金项目(2022110043008507) 

Image Segmentation Based on Federal Style Transfer

  • Received:2022-10-24 Revised:2023-01-14 Online:2023-12-28 Published:2023-12-29

摘要: 针对联邦学习中的非独立同分布问题,提出了一种基于联邦风格迁移的影像分割方法,联邦风格迁移通过共享用户隐私中不敏感的信息,以此生成合成数据进行数据增广,在保证数据重要结构信息不泄露的情况下降低不同用户之间的数据差异。实验结果表明,联邦风格迁移可有效降低肝脏图像分割中各节点之间的非独立同分布现象对联邦模型性能的影响,在统一测试集上的准确率逼近数据集中的训练结果。联邦风格迁移可提高联邦模型的性能,为打破数据孤岛和建立医疗领域的通用分割模型提供可能。

关键词: 图像分割, 深度学习, 风格迁移, 联邦学习

Abstract: In this work, we propose an image segmentation method based on federated style transfer to solve the non-independent and identically distributed (non-IID) problem in federated learning. By sharing style information that is not sensitive to user privacy, this method generates synthetic data for data expansion and reduces data differences between different users while ensuring that important structural information of data is not disclosed. Experiment results show that this method effectively alleviates the influence of non-IID problem among nodes on the performance of the federated model in the liver image segmentation task. Therefore, the proposed method can further improve the performance of federal model, which provides the possibility to break the data island and establish a general model in medical field.

Key words: image segmentation, deep learning, style transfer, federated learning

中图分类号: