| 引用本文: | 马梓博,米悦,张波,张征,吴静云,黄海文,王文东.面向异质性医学图像处理的深度学习算法综述.软件学报,2023,34(10):4870-4915 |
| |
|
| |
|
|
| 本文已被：浏览 4105次 下载 8949次 |
码上扫一扫！ |
|
|
| 面向异质性医学图像处理的深度学习算法综述 |
|
马梓博1,2, 米悦3,4,5, 张波1,2, 张征6, 吴静云7, 黄海文3,4,5, 王文东1,8
|
|
1.网络与交换国家重点实验室 (北京邮电大学), 北京 100876;2.北京邮电大学 计算机学院, 北京 100876;3.北京大学第一医院 泌尿外科, 北京 100034;4.北京大学 泌尿外科研究所, 北京 100034;5.国家泌尿男生殖系肿瘤研究中心, 北京 100034;6.北京邮电大学 现代邮政学院, 北京 100876;7.北京大学第一医院 医学影像科, 北京 100034;8.北京邮电大学 国际学院, 北京 100876
|
|
| 摘要: |
| 近年来深度学习技术在诸多计算机视觉任务上取得了令人瞩目的进步, 也让越来越多的研究者尝试将其应用于医学图像处理领域, 如面向高通量医学图像(CT、MRI)的解剖结构分割等, 旨在为医生提供诊断辅助, 提高其阅片效率. 由于训练医学图像处理的深度学习模型同样需要大量的标注数据, 同一医疗机构的数据往往不能满足需求, 而受设备和采集协议的差异的影响, 不同医疗机构的数据具有很大的异质性, 这导致通过某些医疗机构的数据训练得到模型很难在其他医疗机构的数据上取得可靠的结果. 此外, 不同的医疗数据在患者个体病情阶段的分布上也往往是十分不均匀的, 这同样会降低模型的可靠性. 为了减少数据异质性的影响, 提高模型的泛化能力, 域适应、多站点学习等技术应运而生. 其中域适应技术作为迁移学习中的研究热点, 旨在将源域上学习的知识迁移到未标记的目标域数据上; 多站点学习和数据非独立同分布的联邦学习技术则旨在在多个数据集上学习一个共同的表示, 以提高模型的鲁棒性. 从域适应、多站点学习和数据非独立同分布的联邦学习技术入手, 对近年来的相关方法和相关数据集进行了综述、分类和总结, 为相关研究提供参考. |
| 关键词: 医学图像处理 深度学习 数据异质性 域适应 多站点学习 联邦学习 |
| DOI：10.13328/j.cnki.jos.006680 |
| 分类号: |
| 基金项目:北京市自然科学基金-海淀原始创新联合基金(L182034);国家自然科学基金(61802022,61802027);中央高校基本科研业务费提升科技创新能力行动计划(2019XD-A12);中央高校基本科研业务费专项资金(2020RC07) |
|
| Review on Deep Learning Algorithms for Heterogeneous Medical Image Processing |
|
MA Zi-Bo1,2, MI Yue3,4,5, ZHANG Bo1,2, ZHANG Zheng6, WU Jing-Yun7, HUANG Hai-Wen3,4,5, WANG Wen-Dong1,8
|
|
1.State Key Laboratory of Networking and Switching Technology (Beijing University of Posts and Telecommunications), Beijing 100876, China;2.School of Computer Science, Beijing University of Posts and Telecommunications, Beijing 100876, China;3.The Department of Urology, Peking University First Hospital, Beijing 100034, China;4.The Institute of Urology, Peking University, Beijing 100034, China;5.The National Urological Cancer Center of China, Beijing 100034, China;6.School of Modern Post, Beijing University of Posts and Telecommunications, Beijing 100876, China;7.The Department of Radiology, Peking University First Hospital, Beijing 100034, China;8.International School, Beijing University of Posts and Telecommunications, Beijing 100876, China
|
| Abstract: |
| In recent years, deep learning technology has made remarkable progress in many computer vision tasks. More and more researchers have tried to apply it to medical image processing, such as the segmentation of an atomical structures in high-throughput medical images (CT, MRI), which can improve the efficiency of image reading for doctors. Deep learning models for training medical image processing need a large amount of labeled data, and the data from a separate medical institution can not meet this requirement. Moreover, due to the difference in medical equipment and acquisition protocols, the data from different medical institutions are largely heterogeneous. This results in the difficulty in obtaining reliable results on the data of a certain medical institution with the model trained by data from other medical institutions. In addition, the distribution of different medical data in patients’ disease stages is uneven, thereby reducing the reliability of the model. Technologies including domain adaptation and multi-site learning emerge to reduce the impact of data heterogeneity and improve the generalization ability of the model. As a research hotspot in transfer learning, domain adaptation is intended to transfer knowledge learned from the source domain to data of the unlabeled target domain. Multi-site learning and federated learning with non-independent and identically distributed data aim to improve the robustness of the model by learning a common representation on multiple datasets. This study investigates, analyzes, and summarizes domain adaptation, multi-site learning, and federated learning with non-independent and identically distributed datasets in recent years, providing references for related research. |
| Key words: medical image processing deep learning data heterogeneity domain adaptation multi-site learning federated learning |
|
|
|
|
