AC-UKAN: An enhanced U-KAN framework integrating modulated deformable convolution and mixed pooling

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate medical image segmentation is pivotal for clinical diagnosis. Segmentation targets in medical images differ in appearance, geometric morphology, and contextual organization, which motivates complementary architectural choices rather than a single universal mechanism. We propose AC-UKAN, a depth-specialized extension of U-KAN that assigns modulated deformable convolution (MDConv) to the shallow encoder for input-conditioned geometric sampling, tokenized Kolmogorov–Arnold Network (Tok-KAN) blocks to the low-resolution semantic stages for an alternative nonlinear parameterization, and a mixed pooling module (MPM) to the bottleneck for combined isotropic and anisotropic context aggregation. AC-UKAN was evaluated on the three heterogeneous two-dimensional binary segmentation benchmarks used by U-KAN: BUSI (breast ultrasound), GlaS (gland histopathology), and CVC-ClinicDB (colonoscopic imaging). For a controlled comparison, eight representative segmentation architectures were trained locally under an identical protocol. AC-UKAN achieved an average IoU of 80.78% and DSC of 88.86% with 6.66 M parameters, improving IoU over the U-KAN baseline by up to 4.04 percentage points on the evaluated splits, with its largest gains on BUSI and CVC. Component ablations showed that MDConv and MPM each provided incremental improvements in overlap-based metrics relative to the baseline and to each other’s single-module configuration, whereas their combined effect and boundary-distance behavior varied across datasets. Under the official U-KAN protocol, AC-UKAN therefore offers a favorable accuracy–efficiency trade-off among the evaluated representative architectures, rather than uniform superiority across all three datasets.