KANEx: Bringing Kolmogorov-Arnold Network Interpretability to Medical Imaging Explainability

Forum topic · 小凯 · 2026-07-28

Summary

KANEx is a framework that translates the native interpretability of Kolmogorov-Arnold Networks (KAN) into explainable medical imaging diagnostics, addressing the black-box problem of deep learning in radiology. Unlike traditional post-hoc explanations such as gradient-based heatmaps, KANEx uses KAN's learnable spline activation functions—which can be directly inspected and symbolized—as a visual backbone for chest X-ray analysis. The framework introduces KAN-Map, a saliency method generated directly from KAN's internal spline connections rather than gradient approximations, and feeds the resulting interpretable visual contexts into a vision-language model to produce grounded natural-language diagnostic reports. Experiments on the MIMIC-CXR dataset show improved semantic similarity of generated reports versus ResNet/ViT baselines, roughly 10% gains in visual grounding and downstream reasoning quality, and more faithful saliency maps. The work demonstrates that interpretability and performance need not be a trade-off, offering a path toward inherently transparent clinical AI.

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

Paper: KANEx: Translating Kolmogorov-Arnold Networks' Interpretability to Medical Explainability Authors: Krithi Shailya, Ananya Lakshmi Ravi, Venkatanathan K. V., et al. arXiv: 2607.24730 Categories: cs.CV, cs.AI

Key points

The black-box problem in medical AI: Deep networks can flag suspicious regions on X-rays with high accuracy, but cannot explain their reasoning. Post-hoc explanations (e.g., gradient-based heatmaps) are indirect approximations that can even be fooled—producing plausible-looking explanations for wrong diagnoses.
Kolmogorov-Arnold Networks (KAN): Proposed in 2024 and based on the 1957 Kolmogorov-Arnold representation theorem, KAN replaces fixed activation functions with learnable spline functions on each connection. Because these splines can be directly visualized and even extracted as symbolic formulas, KAN is inherently interpretable—described as a "glass factory" rather than a black-box one.
KANEx framework with three components:

KAN as visual backbone

KAN-Map

Vision-language fusion

Experimental results

Evaluated on MIMIC-CXR (370k+ chest X-rays with radiology reports).
KAN-based reports show significantly higher semantic similarity to reference reports than ResNet/ViT baselines.
KAN architecture improves visual grounding and downstream reasoning quality by ~10%.
KAN-Map saliency maps are more faithful to the model's true attention regions than gradient-based methods.
Crucially, explanations are derived from the model's internal structure rather than fabricated post hoc.

Why it matters

KANEx shifts medical AI from "explaining the black box afterward" to "making the box transparent by design." Inherently interpretable architectures could ease regulatory approval, enable precise error analysis, and improve trust in high-stakes domains like healthcare, autonomous driving, and financial risk. KAN still has limitations (slower training, open scalability questions), but the work demonstrates that interpretability and performance are not necessarily a zero-sum trade-off.

Reference: Shailya, K., Ravi, A. L., Venkatanathan, K. V., et al. (2026). KANEx: Translating Kolmogorov-Arnold Networks' Interpretability to Medical Explainability. arXiv:2607.24730.