Beyond attention heatmaps: How to get better explanations for multiple instance learning models in histopathology.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Multiple instance learning (MIL) is crucial for analyzing gigapixel whole slide images in computational histopathology.
• Heatmaps are commonly used for MIL model validation and biomarker discovery, but their reliability is under-explored.
• Evaluating heatmap quality is essential for trustworthy AI in pathology.

Purpose of the Study:

• Introduce a general framework to assess the quality of MIL heatmaps without additional labels.
• Benchmark six explanation methods across diverse histopathology tasks, MIL architectures, and patch encoders.
• Demonstrate the utility of high-quality MIL heatmaps for biological validation and insight discovery.

Main Methods:

• Developed a novel framework for evaluating MIL heatmap quality.
• Conducted a large-scale benchmark experiment comparing six explanation methods.
• Assessed methods across classification, regression, and survival tasks, and various MIL/encoder architectures.

Main Results:

• Explanation quality is significantly influenced by MIL model architecture and task type.
• Perturbation ('Single'), Layer-wise Relevance Propagation (LRP), and Integrated Gradients (IG) methods outperformed others.
• Attention-based and gradient-based saliency heatmaps often failed to accurately reflect model decision-making.
• Demonstrated correlation of MIL heatmaps with spatial transcriptomics and discovered distinct HPV prediction strategies.

Conclusions:

• Validated MIL heatmaps are critical for reliable model assessment and biological discovery in digital pathology.
• The proposed framework and best-performing methods (Perturbation, LRP, IG) enhance explainability and trustworthiness of MIL models.
• Promotes broader adoption of explainable AI (XAI) for advancing digital pathology research and clinical applications.