多实例学习在医学图像分析中的应用进展  

Application Progress of Multi-instance Learning in Medical Image Analysis

作　　者：谢卓恒 伊鸣[2,3] 黄新瑞 XIE Zhuoheng;YI Ming;HUANG Xinrui(Department of Biophysics,School of Basic Medical Sciences,Peking University,Beijing 100191,China;Neuroscience Research Institute,Peking University,Beijing 100191,China;Key Lab for Neuroscience,the Ministry of Education of China,Key Lab for Neuroscience,the National Health and Family Planning Commission of China,Beijing 100191,China)

机构地区：[1]北京大学基础医学院生物物理学系,北京100191 [2]北京大学神经科学研究所,北京100191 [3]神经科学教育部重点实验室/国家卫生健康委员会神经科学重点实验室,北京100191

出　　处：《集成技术》2025年第2期24-32,共9页Journal of Integration Technology

基　　金：北京大学医学部教育教学研究课题项目(2022YB17);北京市自然科学基金面上项目(4242004);国家蛋白质科学研究(北京)设施北京大学分中心开放课题项目(KF-202402);国家自然科学基金项目(32271053);北京市自然科学基金-海淀原始创新联合基金项目(L222016)。

摘　　要：多实例学习(multiple-instance learning,MIL)是一种弱监督学习方法,近年来广泛应用于医学图像分析领域。本文综述了MIL在全切片图像中的应用进展,详细分析了其在肿瘤检测、亚型分级和生存预测中的作用。MIL在弱监督学习中具有独特优势,可通过引入新机制进行优化和拓展,以适应更多的应用场景。本文首先综述了部分应用广泛或独具优势的MIL模型,并详细介绍了它们的技术特点和具体应用场景;其次,介绍了MIL在多模态医学图像分析中的应用进展和技术进步;最后,总结了MIL目前的研究进展,并展望了其未来发展。Multiple-instance learning(MIL),as a weakly supervised learning method,has been widely applied in the field of medical image analysis in recent years.The paper reviews the progress of MIL applications in whole slide images,with a detailed analysis of its roles in tumor detection,subtype classification,and survival prediction.MIL holds unique advantages in weakly supervised learning,which can be optimized and extended through the introduction of new mechanisms to adapt to a broader range of application scenarios.The paper first reviews some widely used or uniquely advantageous MIL models,elaborating on their technical features and specific application contexts.Secondly,it introduces the application and technology advancements of MIL in multimodal medical image analysis.Finally,the current research progress of MIL is summarized,and its future development prospects are explored.

关 键 词：图像分析 多实例学习 医学图像 机器学习 深度学习 

分 类 号：R319[医药卫生—基础医学]