ABMIL-BiGRU:bidirectional gated recurrent unit attention based multi-instance learning for interpretable prediction of sentinel lymph node metastasis in breast cancer
10.3969/j.issn.1005-202X.2025.02.006
- VernacularTitle:ABMIL-BiGRU:基于双向门控循环注意力多示例学习的乳腺癌淋巴结转移可解释性预测
- Author:
Bo LI
1
;
Yanbin YANG
;
Shuai LI
;
Meiyan LIANG
Author Information
1. 山西省荣军医院物理治疗科,山西 太原 030031
- Publication Type:Journal Article
- Keywords:
breast cancer;
lymph node metastasis;
precise diagnosis;
bidirectional gated recurrent unit;
contextual information;
interpretability;
whole slide image
- From:
Chinese Journal of Medical Physics
2025;42(2):175-183
- CountryChina
- Language:Chinese
-
Abstract:
Aiming at the classification and lesion localization of giga-pixel pathology whole slide images of breast cancer,a bidirectional gated recurrent unit attention based multi-instance learning(ABMIL-BiGRU)model is proposed for interpretable prediction of H&E stained breast cancer lymph node metastasis images.The method uses two orthogonal bidirectional gated recurrent units to establish the long-short distance dependencies between the features in the row and column directions of the image block,thereby embedding the spatial position and context information of the image block,and then quantifies the attention score of each feature representation through attention multi-instance pooling,thereby achieving whole slide image-level feature aggregation and generating interpretable heat maps.The results show that ABMIL-BiGRU model has an average accuracy of 0.918 6 and an AUCof 0.9467 on the breast cancer metastasis dataset,realizing high-precision prediction of whole slide images and localization of regions of interest,and also providing human-interpretable features at the image block level.The proposed model solves the"accuracy-interpretability trade-off"problem to a certain extent,and its superior performance provides a new paradigm for the clinical application of computer-aided diagnosis and intelligent systems.
