AINet: Anchor Instances Learning for Regional Heterogeneity in Whole Slide Image

Tingting Zheng; Hongxun Yao; Kui Jiang; Sicheng Zhao; Yi Xiao

eess.IVFeb 21, 2026

AINet: Anchor Instances Learning for Regional Heterogeneity in Whole Slide Image

Authors: Tingting Zheng, Hongxun Yao, Kui Jiang, Sicheng Zhao, Yi Xiao

Organizations: Harbin Institute of Technology · Tsinghua University · Zhengzhou University

Abstract

Recent advances in multi-instance learning (MIL) have witnessed impressive performance in whole slide image (WSI) analysis. However, the inherent sparsity of tumors and their morphological diversity lead to obvious heterogeneity across regions, posing significant challenges in aggregating high-quality and discriminative representations. To address this, we introduce a novel concept of anchor instance (AI), a compact subset of instances that are representative within their regions (local) and discriminative at the bag (global) level. These AIs act as semantic references to guide interactions across regions, correcting non-discriminative patterns while preserving regional diversity. Specifically, we propose a dual-level anchor mining (DAM) module to select AIs from massive instances, where the most informative AI in each region is extracted by assessing its similarity to both local and global embeddings. Furthermore, to ensure completeness and diversity, we devise an anchor-guided region correction (ARC) module that explores the complementary information from all regions to correct each regional representation. Building upon DAM and ARC, we develop a concise yet effective framework, AINet, which employs a simple predictor and surpasses state-of-the-art methods with substantially fewer FLOPs and parameters. Moreover, both DAM and ARC are modular and can be seamlessly integrated into existing MIL frameworks, consistently improving their performance.

Figures & tables

Figure 1: Comparison results of FLOPs (G), parameters (M), and average accuracy with representative MIL methods on TCGA-ESCA, TCGA-BRCA, and BRACS datasets. AINet delivers superior performance with improved efficiency.

Figure 2: (a) Regional heterogeneity within a WSI originates from tumor sparsity and morphological diversity. (b) Previous MIL methods often neglect this heterogeneity, resulting in suboptimal feature aggregation. (c) The proposed AINet alleviates this limitation via DAM and ARC modules, yielding more discriminative representations and improved predictive accuracy.

Figure 3: Overview of the proposed AINet for WSI classification. Each bag is first divided into multiple spatial regions. The dual-level anchor mining (DAM) module identifies a compact set of anchor instances (AIs), which are then incorporated by the anchor-guided region correction (ARC) module into regional representations to guide the aggregation of heterogeneous features.

Methods		TCGA-BRCA			TCGA-ESCA
Methods		Accuracy	AUC	F1	Accuracy	AUC	F1
ResNet18	ABMIL [ 15 ]	0.862 $\pm$ 0.025	0.882 $\pm$ 0.033	0.915 $\pm$ 0.015	0.815 $\pm$ 0.129	0.882 $\pm$ 0.126	0.852 $\pm$ 0.101
	DSMIL [ 2 ]	0.823 $\pm$ 0.021	0.820 $\pm$ 0.033	0.892 $\pm$ 0.014	0.712 $\pm$ 0.085	0.823 $\pm$ 0.078	0.773 $\pm$ 0.074
	CLAM [ 25 ]	0.865 $\pm$ 0.020	0.890 $\pm$ 0.029	0.917 $\pm$ 0.014	0.821 $\pm$ 0.091	0.916 $\pm$ 0.052	0.844 $\pm$ 0.085
	TransMIL [ 31 ]	0.847 $\pm$ 0.021	0.846 $\pm$ 0.036	0.905 $\pm$ 0.013	0.815 $\pm$ 0.081	0.878 $\pm$ 0.086	0.847 $\pm$ 0.065
	DTFD-MaxM [ 46 ]	0.816 $\pm$ 0.023	0.810 $\pm$ 0.033	0.885 $\pm$ 0.013	0.821 $\pm$ 0.067	0.914 $\pm$ 0.075	0.851 $\pm$ 0.057

Table 1: Quantitative results of AINet compared with representative MIL baselines on TCGA-BRCA and TCGA-ESCA datasets.

Methods	ResNet18			PLIP
Methods	Accuracy	AUC	F1	Accuracy	AUC	F1
ABMIL [ 15 ]	0.691 $\pm$ 0.041	0.816 $\pm$ 0.020	0.604 $\pm$ 0.055	0.754 $\pm$ 0.042	0.875 $\pm$ 0.036	0.705 $\pm$ 0.044
DSMIL [ 2 ]	0.657 $\pm$ 0.026	0.795 $\pm$ 0.021	0.555 $\pm$ 0.016	0.717 $\pm$ 0.044	0.847 $\pm$ 0.039	0.629 $\pm$ 0.041
CLAM [ 25 ]	0.689 $\pm$ 0.036	0.830 $\pm$ 0.028	0.601 $\pm$ 0.024	0.742 $\pm$ 0.054	0.871 $\pm$ 0.023	0.692 $\pm$ 0.043
TransMIL [ 31 ]	0.706 $\pm$ 0.044	0.799 $\pm$ 0.035	0.596 $\pm$ 0.036	0.709 $\pm$ 0.032	0.851 $\pm$ 0.022	0.640 $\pm$ 0.059
DTFD-MaxM [ 46 ]	0.698 $\pm$ 0.030	0.815 $\pm$ 0.062	0.610 $\pm$ 0.044	0.724 $\pm$ 0.054	0.838 $\pm$ 0.063	0.617 $\pm$ 0.075

Table 2: Quantitative results of AINet compared with representative MIL baselines on the BRACS dataset.

Methods	ResNet18			PLIP			FLOPs (G)	#Param. (M)
Methods	Accuracy	AUC	F1	Accuracy	AUC	F1	FLOPs (G)	#Param. (M)
TransMIL [ 31 ]	0.815 $\pm$ 0.081	0.878 $\pm$ 0.086	0.847 $\pm$ 0.065	0.904 $\pm$ 0.062	0.988 $\pm$ 0.018	0.920 $\pm$ 0.051	84.22	0.68
+AINet	0.822 $\pm$ 0.121 (0.7)	0.889 $\pm$ 0.116 (1.1)	0.8438 $\pm$ 0.108 (0.4)	0.916 $\pm$ 0.054 (1.2)	0.993 $\pm$ 0.006 (0.5)	0.923 $\pm$ 0.052 (0.3)	22.30 (61.92)	0.19 (0.49)
ILRA [ 41 ]	0.803 $\pm$ 0.136	0.896 $\pm$ 0.097	0.838 $\pm$ 0.116	0.892 $\pm$ 0.105	0.965 $\pm$ 0.057	0.900 $\pm$ 0.103	123.05	1.85
+AINet	0.866 $\pm$ 0.098 (6.3)	0.915 $\pm$ 0.091 (1.9)	0.882 $\pm$ 0.090 (4.4)	0.905 $\pm$ 0.099 (1.3)	0.967 $\pm$ 0.054 (0.2)	0.918 $\pm$ 0.083 (1.8)	21.86 (101.19)	0.19 (1.66)
WIKG [ 18 ]	0.770 $\pm$ 0.131	0.874 $\pm$ 0.106	0.793 $\pm$ 0.121	0.892 $\pm$ 0.093	0.949 $\pm$ 0.063	0.900 $\pm$ 0.093	49.80	0.43

Table 3: Validation on the TCGA-ESCA dataset. The red denotes performance improvement, while green indicates reduced complexity.

Model	DAM	ACF	ARC	TCGA-ESCA			BRACS
Model	DAM	ACF	ARC	Accuracy	AUC	F1	Accuracy	AUC	F1
Baseline	✗	✗	✗	0.777 $\pm$ 0.134	0.866 $\pm$ 0.095	0.808 $\pm$ 0.114	0.691 $\pm$ 0.056	0.809 $\pm$ 0.054	0.602 $\pm$ 0.073
w DAM	✓	✗	✗	0.809 $\pm$ 0.117	0.886 $\pm$ 0.128	0.841 $\pm$ 0.095	0.713 $\pm$ 0.047	0.835 $\pm$ 0.040	0.594 $\pm$ 0.061
w DAM+MHA	✓	✗	✗	0.839 $\pm$ 0.088	0.935 $\pm$ 0.064	0.867 $\pm$ 0.069	0.719 $\pm$ 0.061	0.845 $\pm$ 0.041	0.624 $\pm$ 0.063
w DAM+ACF	✓	✓	✗	0.840 $\pm$ 0.108	0.938 $\pm$ 0.073	0.868 $\pm$ 0.083	0.732 $\pm$ 0.039	0.844 $\pm$ 0.036	0.631 $\pm$ 0.045
w DAM+ARC (AINet)	✓	✓	✓	0.873 $\pm$ 0.098	0.946 $\pm$ 0.046	0.894 $\pm$ 0.080	0.737 $\pm$ 0.025	0.854 $\pm$ 0.046	0.638 $\pm$ 0.040

Table 4: Ablation study of the key components on the TCGA-ESCA and BRACS datasets with ResNet18 features. The gray highlights our results. We set

L=4

regions, select the top

k=20\%

instances as AIs, and apply a masking ratio of

r=90\%

within each region.

Model	Bag-level	Region-level	TCGA-ESCA			BRACS
Model	Bag-level	Region-level	Accuracy	AUC	F1	Accuracy	AUC	F1
Baseline	✗	✗	0.777 $\pm$ 0.134	0.866 $\pm$ 0.095	0.808 $\pm$ 0.114	0.691 $\pm$ 0.056	0.809 $\pm$ 0.054	0.602 $\pm$ 0.073
w Attention	✗	✗	0.816 $\pm$ 0.156	0.890 $\pm$ 0.109	0.839 $\pm$ 0.131	0.713 $\pm$ 0.043	0.837 $\pm$ 0.041	0.638 $\pm$ 0.054
w Max-pooling	✗	✗	0.808 $\pm$ 0.093	0.918 $\pm$ 0.093	0.842 $\pm$ 0.067	0.681 $\pm$ 0.040	0.820 $\pm$ 0.058	0.593 $\pm$ 0.057
w Bag-similar	✓	✗	0.828 $\pm$ 0.133	0.923 $\pm$ 0.075	0.862 $\pm$ 0.099	0.687 $\pm$ 0.048	0.816 $\pm$ 0.040	0.575 $\pm$ 0.051
w Region-similar	✗	✓	0.834 $\pm$ 0.140	0.920 $\pm$ 0.069	0.869 $\pm$ 0.107	0.728 $\pm$ 0.044	0.835 $\pm$ 0.047	0.619 $\pm$ 0.059

Table 5: Ablation study of the proposed DAM on the TCGA-ESCA and BRACS datasets using ResNet18 features. We set

L=4

regions, select the top

k=20\%

instances as AIs, and apply a masking ratio of

r=90\%

within each region.

Figure 4: Hyperparameter effects on the TCGA-ESCA dataset.

Whole-Slide Image Analysis

Whole-slide image analysis uses AI to interpret gigapixel pathology slides for diagnosis, prognosis, tissue mapping, and report generation.

See all 214 papers

Test-Time Instance Selection for Improved Whole Slide Image Analysis
Aug 14, 2026Quoc Anh Nguyen, Sunhong Park, Jin Tae KwakWhole-Slide Image Analysis Histopathology Image Classification
School of Electrical Engineering, Korea University, Seoul 02841, Korea
ASMIL: Attention-Stabilized Multiple Instance Learning for Whole Slide Imaging
Mar 1, 2026Linfeng Ye, Shayan Mohajer Hamidi, Zhixiang Chi +5Multiple Instance Learning Whole-Slide Image Analysis
University of Toronto · Stanford University · Hokkaido University
ReconMIL: Synergizing Latent Space Reconstruction with Bi-Stream Mamba for Whole Slide Image Analysis
Mar 20, 2026Lubin Gan, Jing Zhang, Heng Zhang +4Histopathology Image Classification Medical Image Localization
USTC, Anhui, China · Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Anhui, China · SCNU, Guangdong, China +2

Aug 14, 2026J/K move · Enter open · S save

Test-Time Instance Selection for Improved Whole Slide Image Analysis

Quoc Anh Nguyen, Sunhong Park, Jin Tae Kwak

School of Electrical Engineering, Korea University, Seoul 02841, Korea

Whole-Slide Image Analysis Histopathology Image Classification Data Selection Multiple Instance Learning

Open paper AskPDF arXiv

Whole Slide Image (WSI) analysis has been widely studied for cancer diagnosis. Conventionally, a gigapixel WSI is divided into small patches and processed by Multiple Instance Learning (MIL) models. However, existing MIL models typically process all patches, many of which contain redundant or non-informative tissue patterns. Although recent approaches have focused on instance selection to identify discriminative patches and reduce redundancy, these selection modules still require additional training. In this work, we propose Test-Time Instance Selection (TTIS), a training-free, plug-and-play framework that selects compact yet representative patches during inference. TTIS further incorporates a multi-view ensemble strategy to integrate distinct facets of tissue morphology, enhancing robustness. Importantly, TTIS can be seamlessly integrated into existing MIL models without retraining or architectural changes, enabling flexible deployment. Extensive evaluations across multiple benchmarks demonstrate that our approach improves or matches baseline MIL performance across a range of classification and subtyping tasks. Our implementation code is available at https://github.com/QuIIL/TTIS

Abstract

Whole-Slide Image Analysis

Explore similar work

Test-Time Instance Selection for Improved Whole Slide Image Analysis

ASMIL: Attention-Stabilized Multiple Instance Learning for Whole Slide Imaging

ReconMIL: Synergizing Latent Space Reconstruction with Bi-Stream Mamba for Whole Slide Image Analysis