UncerKAN-Mamba: A Clinically Robust, Transparent, and Explainable AI Framework for Low-Latency Skin Lesion Segmentation with Deterministic Single-Pass Uncertainty Estimation
Background: Accurate skin lesion segmentation is central to early melanoma detection, yet existing uncertainty estimation methods such as Monte Carlo (MC) Dropout and Deep Ensembles impose heavy computational overhead, and most segmentation architectures offer no insight into where or why predictions may fail. Methods: We present UncerKAN-Mamba, an explainable segmentation architecture integrating an EfficientNet-B4 encoder with Mamba state space model (SSM) blocks inside a UNet++ decoder, augmented by a Kolmogorov–Arnold Network (KAN) uncertainty head. The KAN head uses spline variance across multiple basis degrees as a deterministic, single-pass uncertainty proxy. Explainability is assessed via Grad-CAM, error–uncertainty overlap, and boundary uncertainty profiling. Results: Trained on the ISIC 2018 training partition (n = 1815 of 2594 total images) and externally validated on PH2 (n = 200) and ISIC 2016 (n = 1279), UncerKAN-Mamba achieved Dice = 0.8958, 0.9214, and 0.9360, respectively, achieving segmentation performance statistically comparable to the strongest contemporary baselines on ISIC 2018 and PH2, and ranking second to Attention U-Net on ISIC 2016 (Bonferroni-corrected Wilcoxon and paired bootstrap tests). KAN uncertainty yielded the strongest Pearson correlation with segmentation error (r = 0.674–0.731, best on ISIC 2016)—1.7–2.0× higher than MC Dropout and 1.3–1.4× higher than Deep Ensembles. Quantitative XAI metrics (AUROC = 0.971, PAvPU = 0.554, ECE = 0.0126–0.0419) confirmed strong interpretability and excellent calibration at 36–47 FPS. Conclusions: UncerKAN-Mamba delivers clinically robust, transparent, low-latency skin lesion segmentation suitable for interactive clinical review with deterministic single-pass uncertainty—the first use of KAN spline variance for uncertainty quantification in medical image segmentation.