SpectraNet: a novel model for polyp segmentation leveraging a spectral-guided mixture of functional experts.

Automated and precise polyp segmentation from colonoscopy images is critical for the early diagnosis of colorectal cancer. However, this task is challenged by the ambiguous and low-contrast boundaries of polyps, which often blend with the surrounding mucosa. To address this, we propose SpectraNet, a novel hybrid-domain enhancement network for high-precision polyp segmentation. Our model is built on an encoder-decoder architecture with two core innovations integrated into its skip connections: (1) a Spectral-Guided Boundary Enhancement (SGBE) module that operates in the frequency domain to recover and sharpen indistinct boundary information by enhancing the phase spectrum of features, and (2) a Function-Specialized Mixture-of-Experts (FS-MoE) module that adaptively refines features for diverse polyp morphologies using a set of heterogeneous, function-specific experts. Extensive experiments on our curated PolypSegDataset and two public benchmarks (Kvasir-SEG and CVC-ClinicDB) demonstrate that our method consistently outperforms a wide range of state-of-the-art models. SpectraNet achieves superior performance in key segmentation metrics, and produces qualitatively more accurate segmentation masks with precise boundary definitions.