Sample Hardness Based Gradient Loss for Long-Tailed Cervical Cell Detection

被引:5
作者
Liu, Minmin [1 ,2 ,3 ]
Li, Xuechen [1 ,2 ,3 ,4 ]
Gao, Xiangbo [5 ]
Chen, Junliang [1 ,2 ,3 ]
Shen, Linlin [1 ,2 ,3 ]
Wu, Huisi [1 ,2 ,3 ]
机构
[1] Shenzhen Univ, Comp Vision Inst, Sch Comp Sci & Software Engn, Shenzhen, Peoples R China
[2] Shenzhen Univ, AI Res Ctr Med Image Anal & Diag, Shenzhen, Peoples R China
[3] Shenzhen Univ, Guangdong Key Lab Intelligent Informat Proc, Shenzhen 518060, Peoples R China
[4] Shenzhen Univ, Natl Engn Lab Big Data Syst Comp Technol, Shenzhen 518060, PR, Peoples R China
[5] Univ Calif Irvine, Irvine, CA USA
来源
MEDICAL IMAGE COMPUTING AND COMPUTER ASSISTED INTERVENTION, MICCAI 2022, PT II | 2022年 / 13432卷
基金
中国国家自然科学基金;
关键词
Long-tailed learning; Object detection; Cervical cancer;
D O I
10.1007/978-3-031-16434-7_11
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Due to the difficulty of cancer samples collection and annotation, cervical cancer datasets usually exhibit a long-tailed data distribution. When training a detector to detect the cancer cells in a WSI (Whole Slice Image) image captured from the TCT (Thinprep Cytology Test) specimen, head categories (e.g. normal cells and inflammatory cells) typically have a much larger number of samples than tail categories (e.g. cancer cells). Most existing state-of-the-art long-tailed learning methods in object detection focus on category distribution statistics to solve the problem in the long-tailed scenario, without considering the "hardness" of each sample. To address this problem, in this work we propose a Grad-Libra Loss that leverages the gradients to dynamically calibrate the degree of hardness of each sample for different categories, and rebalance the gradients of positive and negative samples. Our loss can thus help the detector to put more emphasis on those hard samples in both head and tail categories. Extensive experiments on a long-tailed TCT WSI image dataset show that the mainstream detectors, e.g. RepPoints, FCOS, ATSS, YOLOF, etc. trained using our proposed Gradient-Libra Loss, achieved much higher (7.8%) mAP than that trained using crossentropy classification loss.
引用
收藏
页码:109 / 119
页数:11
相关论文
共 25 条
[1]  
Chen K, 2019, Arxiv, DOI arXiv:1906.07155
[2]   You Only Look One-level Feature [J].
Chen, Qiang ;
Wang, Yingming ;
Yang, Tong ;
Zhang, Xiangyu ;
Cheng, Jian ;
Sun, Jian .
2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR 2021, 2021, :13034-13043
[3]   Class-Balanced Loss Based on Effective Number of Samples [J].
Cui, Yin ;
Jia, Menglin ;
Lin, Tsung-Yi ;
Song, Yang ;
Belongie, Serge .
2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2019), 2019, :9260-9269
[4]   The Pascal Visual Object Classes (VOC) Challenge [J].
Everingham, Mark ;
Van Gool, Luc ;
Williams, Christopher K. I. ;
Winn, John ;
Zisserman, Andrew .
INTERNATIONAL JOURNAL OF COMPUTER VISION, 2010, 88 (02) :303-338
[5]  
Goyal P, 2018, Arxiv, DOI arXiv:1706.02677
[6]   LVIS: A Dataset for Large Vocabulary Instance Segmentation [J].
Gupta, Agrim ;
Dollar, Piotr ;
Girshick, Ross .
2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2019), 2019, :5351-5359
[7]   Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning [J].
Han, H ;
Wang, WY ;
Mao, BH .
ADVANCES IN INTELLIGENT COMPUTING, PT 1, PROCEEDINGS, 2005, 3644 :878-887
[8]   Learning from Imbalanced Data [J].
He, Haibo ;
Garcia, Edwardo A. .
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, 2009, 21 (09) :1263-1284
[9]   Learning Deep Representation for Imbalanced Classification [J].
Huang, Chen ;
Li, Yining ;
Loy, Chen Change ;
Tang, Xiaoou .
2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2016, :5375-5384
[10]  
Jantzen J., 2005, P NISIS 2005 NATURE, P1