EfficientMIL: Efficient Linear-Complexity MIL Method for WSI Classification

2025-09-28

AI Summary

Multi-instance learning (MIL) for whole-slide image (WSI) classification suffers from quadratic computational complexity due to self-attention mechanisms, hindering scalability in large-scale pathological analysis. Method: This paper proposes EfficientMIL—a novel framework that introduces linear-complexity state-space models (e.g., Mamba) into WSI analysis for the first time. It replaces global attention with an adaptive patch selection (APS) module to enable efficient and scalable sequence modeling. EfficientMIL unifies linear sequential models—including GRU, LSTM, and Mamba—within a single architecture. Contribution/Results: The framework significantly reduces computational overhead while enhancing discriminative capacity. On TCGA-Lung and CAMELYON16, it achieves AUCs of 0.976/0.990 and accuracies of 0.933/0.975, respectively—outperforming all existing state-of-the-art methods. EfficientMIL establishes a new paradigm for high-accuracy, resource-efficient WSI analysis.

Technology Category

• Machine Learning: Multimodal Learning
• Computer Vision: Multi-modal Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Whole slide images (WSIs) classification represents a fundamental challenge in computational pathology, where multiple instance learning (MIL) has emerged as the dominant paradigm. Current state-of-the-art (SOTA) MIL methods rely on attention mechanisms, achieving good performance but requiring substantial computational resources due to quadratic complexity when processing hundreds of thousands of patches. To address this computational bottleneck, we introduce EfficientMIL, a novel linear-complexity MIL approach for WSIs classification with the patches selection module Adaptive Patch Selector (APS) that we designed, replacing the quadratic-complexity self-attention mechanisms in Transformer-based MIL methods with efficient sequence models including RNN-based GRU, LSTM, and State Space Model (SSM) Mamba. EfficientMIL achieves significant computational efficiency improvements while outperforming other MIL methods across multiple histopathology datasets. On TCGA-Lung dataset, EfficientMIL-Mamba achieved AUC of 0.976 and accuracy of 0.933, while on CAMELYON16 dataset, EfficientMIL-GRU achieved AUC of 0.990 and accuracy of 0.975, surpassing previous state-of-the-art methods. Extensive experiments demonstrate that APS is also more effective for patches selection than conventional selection strategies.

Problem

• Reducing computational complexity in whole slide image classification
• Replacing quadratic-complexity attention with linear-complexity sequence models
• Improving patch selection efficiency while maintaining classification performance

Innovation

• Replaces quadratic self-attention with linear-complexity sequence models
• Introduces Adaptive Patch Selector for efficient patch selection
• Uses RNN, LSTM and Mamba models for WSI classification