Polyp Segmentation by Dual-Domain Reasoning: Fuzzy Spatial Control and Frequency Selection

ICLR 2026 Conference Submission18220 Authors

Keywords: Polyp Segmentation, Fuzzy Learning, Frequency Learning, Mamba

Abstract: Colorectal cancer (CRC) screening relies on accurate polyp segmentation, yet subtle appearance differences and ambiguous boundaries in colonoscopy images make this task challenging. To overcome these limitations, we propose FSFMamba, a dual-domain fusion network that jointly models boundary uncertainty and frequency structure to improve delineation. In the spatial domain, a Fuzzy Spatial Control Mechanism (FSCM) instantiates an interval type-2 membership to localize uncertainty at boundary bands while preserving stability in homogeneous regions. In the spectral domain, a Frequency Adaptive Selection Mechanism (FASM) performs octave-wise spectral decomposition and applies learnable band-wise weighting to emphasize task-relevant subbands and suppress spurious responses. The two streams are fused by a Mamba-based state-space block that enables long-range, low-latency interactions and pre-norm residual refinement for stable optimization. Extensive experiments show FSFMamba consistently outperforms recent baselines with sharper boundaries, fewer false positives, and strong robustness.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18220