Probabilistic smooth attention for deep multiple instance learning in medical imaging

2025-07-20

AI Summary

Existing deep multiple-instance learning (MIL) methods for medical image classification suffer from inadequate modeling of instance-level contribution uncertainty and limited interpretability. To address this, we propose a probabilistic deep MIL framework that explicitly models attention as a probability distribution over instances—rather than deterministic weights—thereby quantifying uncertainty in lesion localization. Our architecture jointly incorporates local neighborhood modeling and global dependency capture, and introduces a smoothness regularization term to enhance robustness in probability distribution estimation. Evaluated on three public medical imaging datasets, our method achieves statistically significant improvements over 11 state-of-the-art baselines in AUC and F1-score. Beyond superior classification performance, it generates interpretable lesion uncertainty maps, offering clinicians quantitative insights into model confidence and thereby improving reliability and clinical trustworthiness.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Computer Vision: Multi-modal Vision

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

The Multiple Instance Learning (MIL) paradigm is attracting plenty of attention in medical imaging classification, where labeled data is scarce. MIL methods cast medical images as bags of instances (e.g. patches in whole slide images, or slices in CT scans), and only bag labels are required for training. Deep MIL approaches have obtained promising results by aggregating instance-level representations via an attention mechanism to compute the bag-level prediction. These methods typically capture both local interactions among adjacent instances and global, long-range dependencies through various mechanisms. However, they treat attention values deterministically, potentially overlooking uncertainty in the contribution of individual instances. In this work we propose a novel probabilistic framework that estimates a probability distribution over the attention values, and accounts for both global and local interactions. In a comprehensive evaluation involving {color{review} eleven} state-of-the-art baselines and three medical datasets, we show that our approach achieves top predictive performance in different metrics. Moreover, the probabilistic treatment of the attention provides uncertainty maps that are interpretable in terms of illness localization.

Problem

• Addresses uncertainty in attention values for medical imaging classification
• Improves interpretability of illness localization via probabilistic attention
• Enhances deep MIL performance with global and local interactions

Innovation

• Probabilistic framework for attention values
• Combines global and local interactions
• Generates interpretable uncertainty maps