KANEx: Translating Kolmogorov-Arnold Networks' Interpretability to Medical Explainability

TL;DR

KANEx integrates Kolmogorov-Arnold Networks' mathematically interpretable architecture with vision-language models to generate more trustworthy medical explanations. It introduces KAN-Map, a heatmap method derived from KAN's symbolic components, improving visual localization and reasoning fidelity. The framework enhances explainability without relying on gradient approximations.

Why It Matters

In clinical settings, trust in AI depends on explainability. By grounding visual explanations in interpretable mathematical functions, KANEx improves both the accuracy of saliency maps and the quality of natural language reasoning, making AI systems more reliable for medical decision support.

Key Findings

• KAN-based models outperform ResNet and ViT baselines in semantic similarity and faithfulness of textual explanations on MIMIC-CXR.
• KAN-Map produces more accurate saliency maps by directly leveraging KAN's spline-based functional units instead of gradient-based approximations.
• Integrating KANs with VLMs improves visual localization and downstream reasoning quality by 10% compared to standard black-box models.