U‐ KAN for Multi‐Nuclei Segmentation Using an Adaptive Sliding Window Approach

ABSTRACT Accurate segmentation of nuclei in histopathological images is critical for improving diagnostic precision and advancing computational pathology. Deep learning models employed for this task must effectively handle structural variability while offering transparent and interpretable predictions to ensure clinical reliability. In this study, we investigate the integration of Kolmogorov–Arnold Networks (KANs) into the widely adopted U‐Net architecture, forming a novel hybrid model referred to as U‐KAN. To the best of our knowledge, we are the first to explore the application of U‐KAN for multi‐class nuclei segmentation on the challenging MoNuSAC2020 dataset, leveraging an adaptive sliding window strategy. Our results demonstrate that U‐KAN achieves a 17.9% improvement in Dice coefficient (Dice Similarity Coefficient, DSC) (0.976) and a 25.7% increase in IoU (Intersection over Union) (0.954) compared to baseline method (U‐Net), while also delivering enhanced model interpretability. Gradient‐based explanation techniques further confirm that U‐KAN produces anatomically plausible predictions, with strong attention to nuclear boundaries. These findings suggest that symbolic‐connectionist hybrids like U‐KAN can meaningfully advance automated histopathological image analysis.