Pathological image classification model based on pseudo-bag strategy and feature adjustment
10.3969/j.issn.1005-202X.2025.06.010
- VernacularTitle:基于伪包策略和特征调整的病理图像分类模型
- Author:
Jinling CHEN
1
;
Yanlin SU
;
Zhouwei TANG
;
Jihong WEI
;
Qi KE
;
Yuzhu JI
;
Ziqing GAO
Author Information
1. 西南石油大学电气信息学院,四川 成都 610500
- Publication Type:Journal Article
- Keywords:
whole-slide pathological image;
pseudo-bag strategy;
selective feature fusion method;
confounding factor;
classification accuracy
- From:
Chinese Journal of Medical Physics
2025;42(6):775-783
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a classification model based on a pseudo-bag strategy and feature adjustment for whole slide imaging in pathology.Methods A pseudo-bag generator was constructed to divide a parent bag into 3 pseudo-bags for increasing the number of training bags.Then,a pseudo-bag learning method based on Nystr?m-based algorithm for approximating self-attention and a selective feature fusion method were employed to process the pseudo-bags.Specifically,the pseudo-bag learning method based on Nystr?m-based algorithm for approximating self-attention reduced computational complexity through an improved multi-head self-attention mechanism while deeply extracting instance features to obtain pseudo-bag classification predictions,thereby enhancing pseudo-bag classification accuracy;and the selective feature fusion method refined pseudo-bag features by filtering and extracting relevant instances.Finally,the model adjusted bag features by extracting confounding factors to avoid interference from irrelevant information and further improve classification accuracy.Results The proposed model was evaluated on two datasets(CAMELYON-16 and TCGA-NSCLC)and compared with 10 other methods,and the results demonstrated that the proposed model achieved the best performance.The proposed method reached an accuracy of 0.943 on the CAMELYON-16 dataset and 0.906 on the TCGA-NSCLC dataset.Conclusion The proposed model can significantly improve the accuracy of whole-slide pathological image classification by effectively mitigating the overfitting and avoiding interference from irrelevant information.
