From slides (through tiles) to pixels: an explainability framework for weakly supervised models in pre-clinical pathology

Weakly supervised AI can work with realistic data constraints: The framework successfully trains deep learning models on whole slide images using only slide-level labels (diagnoses) rather than requiring time-consuming pixel-level annotations by pathologists, addressing a practical bottleneck in digital pathology.

Explainability methods need validation against expert knowledge: The researchers quantitatively validated their explanations by comparing AI-generated heatmaps against pathologists' manual annotations and segmentation models, finding meaningful correlation between where the AI focused and actual tissue lesions—demonstrating that explainability claims should be empirically tested, not assumed.

Explanations show stability but imperfect alignment with ground truth: While the AI's explanations were stable across input variations and improved with model performance, they did not perfectly match pathologist annotations, suggesting that explainability heatmaps are useful guides but should not be treated as definitive diagnostic maps without expert review.

Context matters for AI interpretability in specialized domains: The framework was specifically designed for the unique challenges of pathology (large image sizes, sparse annotations, regulatory requirements), illustrating that responsible AI implementation requires domain-specific approaches rather than one-size-fits-all explainability solutions.