Paper Information
Title: 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
The Black-Box Problem in Medical AI
Deep neural networks can flag suspicious regions in X-rays—sometimes more accurately than human clinicians—but they cannot explain *why*. In medicine this is critical: a false negative can delay cancer treatment, and a false positive can lead to unnecessary biopsies. Doctors must be accountable, patients deserve reasons, and even a 99%-accurate model is hard to accept clinically if its reasoning is opaque.
Common post-hoc explanations (e.g., gradient-based heatmaps) are indirect approximations—like asking a second student to explain why the first student answered a certain way. Worse, such explanations can be fooled: adversarial attacks can produce wrong diagnoses accompanied by plausible-looking heatmaps.
KAN: A Learnable, Transparent Architecture
Kolmogorov-Arnold Networks (KAN), introduced in 2024, build on the 1957 Kolmogorov-Arnold theorem: any multivariate continuous function can be represented as a composition of univariate functions. Instead of fixed activations (ReLU, Sigmoid) with weighted edges, every connection in a KAN is a learnable spline function whose shape adapts during training.
Because these splines can be directly visualized and even extracted as symbolic formulas, KAN's entire computation is transparent by construction—a "glass factory" rather than a black box.
The KANEx Framework
KANEx combines three components:
1. KAN as visual backbone – replaces CNN/ViT encoders for chest X-rays; each spline connection learns an inspectable visual feature transform (edges, textures, tissue structure). 2. KAN-Map – a saliency method generated directly from the KAN's internal spline responses, not via gradient approximation, producing heatmaps that faithfully reflect the model's actual focus. 3. Vision-language fusion – grounded, interpretable visual contexts are fed to a vision-language model (VLM), so generated reports rest on explainable visual evidence rather than black-box guesses.
Results
Experiments on MIMIC-CXR (370k+ images with radiology reports) show:
- Significantly higher semantic similarity between generated and ground-truth reports versus ResNet/ViT baselines.
- A 10% improvement in visual grounding and downstream reasoning quality with KAN architectures.
- Saliency maps more faithful to the model's true attention than gradient-based methods.
- Explanations derived directly from model internals, fundamentally addressing the trustworthiness problem of post-hoc methods.
Why It Matters
KANEx shifts medical AI from "explaining the black box afterward" to "building a transparent box from the start." If safety-critical AI (healthcare, autonomous driving, finance) were inherently interpretable, regulatory approval, error analysis, and public trust could all improve. KANs still train more slowly and face scaling questions, but KANEx demonstrates that interpretability and performance are not necessarily a zero-sum trade-off.
> "If you can't explain it to a six-year-old, you don't understand it yourself." — Richard Feynman
---
Reference: Shailya, K., Ravi, A. L., Venkatanathan, K. V., et al. KANEx: Translating Kolmogorov-Arnold Networks' Interpretability to Medical Explainability. arXiv:2607.24730.