计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (20): 42-52.DOI: 10.3778/j.issn.1002-8331.2106-0103

• 热点与综述 • 上一篇    下一篇

迁移学习在医学图像分析中的应用研究综述

黎英   

  1. 南宁师范大学 广西高校科学计算与智能信息处理重点实验室,南宁 530001
  • 出版日期:2021-10-15 发布日期:2021-10-21

Review of Application of Transfer Learning in Medical Image Analysis

LI Ying   

  1. Guangxi Colleges and Universities Key Laboratory of Scientific Computing and Intelligent Information Processing, Nanning Normal University, Nanning 530001, China
  • Online:2021-10-15 Published:2021-10-21

摘要:

迁移学习是机器学习中一种新的学习范式,它可以克服深度学习需要大量样本的缺陷,能够解决医学图像分析中数据集较小导致模型不准确的问题,因而成为继深度学习之后在医学图像分析领域的研究热点。对迁移学习进行概要阐述,按照目前医学图像分析中应用的主要迁移学习方法,即基于数据的迁移学习、基于模型的迁移学习、对抗式迁移学习和混合迁移学习,对医学图像分析领域的重要文献进行整理和归纳,分析每种迁移学习的机制、适用范围、应用情况和优缺点,再对这几种迁移学习方法进行总结、分析及比较。针对研究现状的不足指出该领域的研究发展趋势,为迁移学习在该领域的深入研究提供参考。

关键词: 迁移学习, 医学图像, 机器学习, 学习范式

Abstract:

Transfer learning is a new learning paradigm in machine learning. It can overcome the defect that deep learning needs a large number of samples, and can solve the problem of inaccurate model caused by small data set in medical image analysis. Therefore, it has become a research hot spot in the field of medical image analysis after deep learning. Firstly, this paper briefly describes the transfer learning, and then according to the main transfer learning methods currently applied in medical image analysis, namely data-based transfer learning, model-based transfer learning, adversarial transfer learning and mixed transfer learning, sorts out and summarizes the important literature in the field of medical image analysis, and analyzes the mechanism, scope of application, application, advantages and disadvantages of each transfer learning method. Then, the paper summarizes, analyzes and compares these transfer learning methods. Finally, the paper points out the development trend of the research in this field according to the deficiency of the research status, and provides a reference for the further research of transfer learning in this field.

Key words: transfer learning, medical images, machine learning, learning paradigm