qZACH-ViT: Quantization-Aware Intrinsic Explanations with Recursive Attribution-Stabilized Optimization

Quantization preserves both accuracy and interpretability: The research demonstrates that compressing medical image AI models to use smaller data types (INT8) maintains nearly perfect agreement with full-precision versions (99.98% prediction match) while reducing model size by 70%, suggesting that efficiency and trustworthiness don't have to be sacrificed for one another.

Explainability can be built into the compression process: Rather than treating model compression and interpretability as separate concerns, the authors' RASO method actively aligns the model's decision-making process with its explanations during compression, achieving a mean gain of 0.0368 in performance metrics—showing that responsible AI design choices can improve rather than hinder model performance.

Explanations remain visually consistent after compression: Across 3,600 comparisons, the visual explanations (attribution maps) showing which image regions the model uses for diagnosis maintained 99.9955% similarity after compression, with 96.92% overlap in the most important regions, indicating that doctors would see essentially identical evidence regardless of which version they deploy.

Practical deployment becomes feasible for resource-limited settings: The compressed models run 1.41× to 2.39× faster on standard computers while being 70% smaller, making interpretable medical AI accessible in clinics or regions with limited computational infrastructure—a key consideration for equitable healthcare AI deployment.