文献类型：期刊文献

英文题名：Improving multi-instance learning with hierarchical attention and frequency-domain hard sample distillation

作者：Xiao, Ting[1,2];Sun, Minqian[2];Xia, Yiqing[2];Yang, Hai[2];Wang, Zhe[1,2];Liu, Peng[3]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200030, Peoples R China;[3]Harbin Inst Technol, Sch Comp Sci & Technol, Harbin 150001, Peoples R China

年份：2026

卷号：347

外文期刊名：KNOWLEDGE-BASED SYSTEMS

收录：;EI(收录号：20262220814710);WOS:【SCI-EXPANDED(收录号:WOS:001783646100001)】；

基金：This work was supported by the National Natural Science Foundation of China (No. 62306115 and No. 62476087) and the independent research project of the State Key Laboratory of Spatial Intelligent Manipulation Technology.

语种：英文

外文关键词：Multiple instance learning; Histological whole slide image; Hard sample mining; Attention regularization

摘要：Weakly supervised multiple instance learning (MIL) has gained widespread employment in whole slide image (WSI) classification, with most methods favoring the bag-based approach due to its good performance. Most approaches incorporate an attention mechanism to generate bag-level representations. However, due to a significant imbalance between the number of positive and negative instances within positive bags, attention-based MIL methods still have a cascade of issues: (1) highly imbalanced instances lead to attention over-concentration; (2) attention over-concentration leads to difficulty in balancing salient and hard instances, and results in misguided attention to irrelevant patterns. This paper proposes a novel MIL framework, named HAFD-MIL, which incorporates hierarchical attention and frequency-domain hard sample distillation. This framework addresses these challenges by jointly considering sample categories and sample difficulty. Specifically, from the perspective of sample categories, we first classify instances within each pseudo-bag as "trend-to-positive", "trend-to-negative", or "weak negative" based on negative instance prototypes clustered from negative bags. Then we propose a graded attention distillation module to process the above-classified instances separately to reduce the influence of negative instances on positive instances, along with a novel dynamic label attention loss to prevent attention over-concentration. From the perspective of sample difficulty, we propose a spectrum attention distillation module designed to extract information from hard samples and a redundant feature filter module to minimize interference from irrelevant information typically involved in the attention mechanism. Experimental results on two WSI datasets with three pre-trained backbones demonstrate the superiority of our HAFD-MIL framework and exhibit a wider range of attention areas.