Yiwei Li, Huanhuan Zhang, Huiping Yao, Rong Xu, Haishan Huang, Suhua Xue, Leijun Shi, Xiujun Yang, Pengchen Liang and Tingting Li
Achieving strong medical image segmentation under tight computational budgets remains challenging: CNNs excel at local detail but struggle with long-range context, while Transformer variants are accurate yet costly. We address this tension with ProtoKAN-F, a dual-frequency, prototype-driven framework that couples frequency-aware encoding with lightweight token reasoning. The architecture adopts a two-stage encoder–decoder design. A compact convolutional stem extracts shallow spatial/texture cues, whose features are decomposed into low- and high-frequency streams via either discrete wavelet transform or learnable low/high-pass filters. Each stream is processed by tokenized Kolmogorov–Arnold Network (Tok-KAN) blocks, where learnable activations (KANLinear) interleave with depthwise convolutions to model nonlinear token interactions while preserving spatial structure. To better capture fine boundaries without sacrificing receptive field, the high-frequency path incorporates a frequency-adaptive dilated convolution whose dilation rate is dynamically modulated by channel-wise spectral energy; residual normalization stabilizes training. A gated cross-path fusion merges complementary cues into a unified, frequency-aware representation. The decoder introduces a prototype-based head that treats learnable queries as latent structural prototypes. Prototypes are formed by similarity matching with the fused token map and top-k token aggregation, then refined via element-wise interactions and a projection to produce the final mask—avoiding heavy convolutional refinement while retaining accuracy. Comprehensive experiments on three public benchmarks and a private pediatric neuroblastoma segmentation dataset show that ProtoKAN-F delivers higher accuracy and stronger boundary preservation at lower parameter counts and FLOPs, yielding faster inference than strong CNN/Transformer baselines.