FreqMamba-Net: a frequency-enhanced state space model for fine-grained mucosal lesion segmentation in colonoscopy

Accurate delineation of inflammatory lesions in colonoscopy is essential for the objective assessment of inflammatory bowel disease (IBD), yet remains challenging due to the diffuse distribution, subtle textural alterations, and ill-defined boundaries of mucosal inflammation. Conventional convolutional and attention-based segmentation models primarily rely on spatial morphology and struggle to capture the fine-grained patterns characteristic of IBD lesions.

Here, we introduce FreqMamba-Net, a dual-stream segmentation framework that jointly models global spatial context and frequency-domain texture information. The spatial pathway leverages state space-based visual blocks to efficiently capture long-range dependencies and contiguous inflammatory regions, while a complementary frequency pathway enhances high-frequency structural cues associated with subtle mucosal abnormalities. These representations are adaptively integrated through a cross-modality gating mechanism, followed by boundary-aware refinement to improve lesion margin delineation. We evaluated the proposed method on a clinically challenging colonoscopy dataset designed to reflect real-world imaging conditions.

FreqMamba-Net consistently outperformed established convolutional and transformer-based models in segmentation accuracy and boundary consistency, while demonstrating strong correlation with clinical severity proxies and robust generalization across varied imaging scenarios.

By unifying efficient global modeling with explicit texture enhancement, FreqMamba-Net enables reliable identification of diffuse inflammatory lesions that are difficult to detect using conventional approaches. This framework provides a scalable foundation for developing quantitative, objective tools to support IBD assessment in both clinical practice and therapeutic trials.