Related Experiment Video
Updated: Jul 6, 2026
13:51
Cross-Modal Multivariate Pattern Analysis
Published on: November 9, 2011
AttriMIL: Revisiting attention-based multiple instance learning for whole-slide pathological image classification
Linghan Cai1, Shenjin Huang2, Ye Zhang1
1School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, 518055, China.
Medical Image Analysis
|May 17, 2025
Summary
AttriMIL enhances whole-slide pathological image analysis by introducing attribute-aware multiple instance learning (MIL). This framework improves disease classification and region localization by better differentiating tissue instances.
Area of Science:
- Computational pathology
- Artificial intelligence in medicine
- Digital pathology
Background:
- Multiple instance learning (MIL) is crucial for whole-slide pathological image (WSI) analysis, especially with slide-level labels.
- Attention-based MIL models advance weakly supervised WSI classification but struggle with instance differentiation, potentially degrading performance.
- Differentiating between instances is key for accurate tissue identification and classification in WSI analysis.
Purpose of the Study:
- To introduce AttriMIL, an attribute-aware multiple instance learning framework to address limitations in current MIL approaches for WSI analysis.
- To enhance the differentiation of instances within gigapixel-resolution pathological images.
- To improve the accuracy and robustness of weakly supervised WSI classification and disease-positive region localization.
Main Methods:
- Developed a multi-branch attribute scoring mechanism to quantify pathological attributes of individual instances.
- Introduced region-wise and slide-wise attribute constraints to model instance correlations dynamically during training.
- Implemented a pathology adaptive learning technique to optimize pre-trained feature extractors for task-specific feature extraction.
Main Results:
- AttriMIL consistently outperformed state-of-the-art methods across five public datasets.
- Demonstrated superior performance in bag classification accuracy, generalization ability, and disease-positive region localization.
- The attribute constraints effectively encouraged the network to capture spatial patterns and semantic similarities, improving sensitivity to challenging instances.
Conclusions:
- AttriMIL provides a robust framework for attribute-aware multiple instance learning in WSI analysis.
- The proposed attribute constraints and adaptive learning technique significantly enhance classification and localization performance.
- AttriMIL offers a promising advancement for computational pathology, aiding clinical diagnosis through improved WSI analysis.
