Abstract
Abstract Accurate white blood cell (WBC) segmentation is essential for automated hematological analysis; however, performance often degrades under staining variability and limited annotated data. We propose CA-KANet, a Kolmogorov–Arnold UNet architecture with channel-aware encoding that integrates three-dimensional convolutional feature extraction with a spline-based Kolmogorov–Arnold Network (KAN) refinement module. By interpreting RGB channels as a structured spectral axis, the encoder captures localized spatial–chromatic interactions and preserves inter-channel feature coupling across hierarchical representations, while the KAN bottleneck enhances nonlinear feature transformation in a compact parameterization. Extensive evaluation on three publicly WBC datasets (LISC, KRD, and Raabin-WBC) demonstrates that CA-KANet achieves competitive and consistent performance across diverse imaging conditions. Ablation studies show the complementary effects of channel-aware encoding and HKAN-based nonlinear refinement. Efficiency analysis suggests a reasonable accuracy-complexity trade-off, though improvements remain incremental. Overall, the results suggest that incorporating spectral–spatial feature coupling with nonlinear refinement can provide a stable and effective design for color microscopy segmentation under heterogeneous imaging conditions.