Probabilistic Attention Based on Gaussian Processes for Deep Multiple Instance Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Multiple Instance Learning (MIL) is a weakly supervised learning method reducing annotation effort, valuable in data-scarce fields like medicine.
• Current deep learning MIL methods achieve high accuracy but lack uncertainty estimation, a critical feature for high-stakes applications.
• Deterministic models are prone to overfitting, especially with limited data, hindering their reliability in medical diagnostics.

Purpose of the Study:

• To introduce a novel probabilistic attention mechanism for deep MIL, named attention Gaussian process (AGP).
• To enable accurate bag-level predictions, instance-level explainability, and provide crucial uncertainty estimations.
• To enhance the robustness and reliability of MIL models, particularly in medical applications with limited annotated data.

Main Methods:

• Developed the attention Gaussian process (AGP), a probabilistic attention mechanism integrating Gaussian Processes (GPs) into deep MIL.
• Trained the AGP model end-to-end, enabling simultaneous learning of attention and predictions.
• Validated the model on synthetic datasets (MNIST, CIFAR-10) and three real-world cancer detection tasks.

Main Results:

• AGP demonstrated superior performance compared to state-of-the-art deterministic deep learning MIL approaches.
• The model achieved strong results even on a small dataset (<100 labels) and showed better generalization on an external test set.
• Predictive uncertainty from AGP was shown to correlate with prediction errors, serving as a reliable indicator of model confidence.

Conclusions:

• The AGP model provides accurate predictions and valuable uncertainty quantification for deep MIL, outperforming existing methods.
• Its probabilistic nature enhances robustness against overfitting and improves reliability, especially in medical contexts.
• AGP offers a promising approach for explainable and trustworthy AI in medical image analysis and other data-limited domains.