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

Imagine you're a clinician relying on AI to interpret chest X-rays. You want to trust the model's decisions, but often, it feels like a black box — you get a diagnosis, but no clear reasoning behind it. This lack of transparency can lead to skepticism about the AI's reliability, especially in critical medical settings. Current approaches often pair visual classifiers with language models to generate explanations, but they don't really solve the problem of understanding how the visual model arrived at its conclusions. This is what's called opacity in AI systems, which can undermine clinician trust and patient care. To address this, researchers have turned to Kolmogorov-Arnold Networks (KANs), which use spline-based components that are inherently interpretable. By leveraging the transparency of KANs, they developed KANEx, a framework that grounds the reasoning of Vision-Language Models (VLMs) in these interpretable units. This approach not only enhances the linguistic explanations provided by the AI but also introduces KAN-Map, a new method for generating heatmaps directly from KAN models, improving the fidelity of visual attributions. Benchmarked on the MIMIC-CXR dataset, KAN architectures showed a 10% improvement in visual localization and reasoning quality, suggesting that grounding explanations in mathematically interpretable components is crucial for building trustworthy medical AI systems.

The introduction of KANs for interpretability in medical AI is a meaningful extension.

The benchmarks on the MIMIC-CXR dataset provide solid evidence of improved performance.