网站首页            期刊简介             编委会             投稿指南             期刊订阅             下载中心             在线留言            联系我们             English
  2025年4月5日 星期六  
文章快速检索
中国生物医学工程学报  2024, Vol. 43 Issue (6): 652-661    DOI: 10.3969/j.issn.0258-8021.2024.06.002
  论著 本期目录 | 过刊浏览 | 高级检索 |
基于聚类的多实例学习全视野数字切片分类
钟海勤, 赵程, 雷柏英*, 汪天富*
深圳大学医学部生物医学工程学院,广东省生物医学信息检测和超声成像重点实验室,广东 深圳 518060
Cluster-Based Multiple Instance Learning for Whole Slide Image Classification
Zhong Haiqin, Zhao Cheng, Lei Baiying*, Wang Tianfu*
Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging,School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, Shenzhen 518060, Guangdong, China
全文: PDF (4459 KB)   HTML (1 KB) 
输出: BibTeX | EndNote (RIS)      
摘要 病理图像是检验癌症的金标准,对病理图像,尤其是全视野数字切片(WSI),进行快速、准确地分类有助于辅助医生对患者进行个性化治疗和预后评估。近年来,多实例学习(MIL)在WSI分类中发挥着越来越重要的作用。然而,由于WSI的数量有限,且阳性区域占比较低,现有的基于注意力机制的MIL方法可能会导致过拟合,从而影响分类的性能。为了解决这个问题,本研究提出一种新的基于聚类的MIL分类方法。具体地说,为了增加包的数量,让网络关注更多的阳性实例,将每个包划分为多个伪包;然后,为了解决在伪包划分过程中容易出现一个伪包全是阴性实例,导致产生噪声的现象,提出一种新的基于聚类的伪包划分方法;最后,为了获得更加精准的分类结果,将学习到的伪包级特征进行二次学习,得到最终的包级特征,并实现最终的WSI分类。在Camelyon16和TCGA-Lung数据集上进行实验,分别有399张WSI和1 038张WSI,分类准确率分别为90.69%和86.54%,F1-评分分别为90.20%和86.52%。实验结果,表明所提出的方法可有效应用于WSI分类中。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
钟海勤
赵程
雷柏英
汪天富
关键词 全视野数字切片多实例学习分类聚类伪包    
Abstract:Pathological images are the gold standard for cancer examination. Fast and accurate classification of pathological images, especially whole slide images (WSI), helps medical doctors provide personalized treatment and prognosis assessment for patients. In recent years, multiple instance learning (MIL) has played an increasingly important role in WSI classification. However, due to the limited number of WSIs and the low proportion of positive areas, the existing MIL method based on attention mechanism may lead to overfitting, thus affecting the classification performance. To solve this problem, we proposed a new clustering-based MIL classification method. Specifically, this method divided each bag into multiple pseudo bags to increase the number of packages and let the network pay attention to more positive instances. Then, to solve the problem that a pseudo-bag is easy to be full of negative instances in the pseudo-bag division process, resulting in noise, this paper proposed a new pseudo-bag division method based on clustering. Finally, to obtain more accurate classification results, we conducted secondary learning on the learned pseudo-bag-level features to get the final bag-level features and achieve the final WSI classification. We conducted experiments on the Camelyon16 and TCGA-Lung datasets, which have 399 and 1 038 WSIs, respectively, with classification accuracies of 90.69% and 86.54%, and F1-scores of 90.20% and 86.52%. The experimental results showed that the proposed method could be appled to WSI classification effectively.
Key wordswhole slide image(WSI)    multiple instance learning(MIL)    classification    clustering    pseudo bag
收稿日期: 2024-04-02     
PACS:  R318  
基金资助:国家自然科学基金(62171312 , 62301329), 广东省区域联合基金(2022A1515110704)
通讯作者: *E-mail: leiby@szu.edu.cn;tfwang@szu.edu.cn   
引用本文:   
钟海勤, 赵程, 雷柏英, 汪天富. 基于聚类的多实例学习全视野数字切片分类[J]. 中国生物医学工程学报, 2024, 43(6): 652-661.
Zhong Haiqin, Zhao Cheng, Lei Baiying, Wang Tianfu. Cluster-Based Multiple Instance Learning for Whole Slide Image Classification. Chinese Journal of Biomedical Engineering, 2024, 43(6): 652-661.
链接本文:  
http://cjbme.csbme.org/CN/10.3969/j.issn.0258-8021.2024.06.002     或     http://cjbme.csbme.org/CN/Y2024/V43/I6/652
版权所有 © 2015 《中国生物医学工程学报》编辑部
本系统由北京玛格泰克科技发展有限公司设计开发