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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate colonoscopic delineation of inflammatory bowel disease (IBD) lesions is crucial but challenging due to subtle textural changes and ill-defined boundaries.
• Current segmentation models struggle with the diffuse nature and fine-grained patterns of mucosal inflammation in IBD.

Purpose of the Study:

• To introduce FreqMamba-Net, a dual-stream framework for enhanced segmentation of IBD lesions in colonoscopy.
• To improve the objective assessment of IBD by accurately identifying inflammatory regions and their boundaries.

Main Methods:

• Developed FreqMamba-Net, a dual-stream segmentation framework integrating spatial context and frequency-domain texture.
• Employed state space-based visual blocks for global context and a frequency pathway for high-frequency structural cues.
• Utilized a cross-modality gating mechanism and boundary-aware refinement for adaptive integration and improved margin delineation.

Main Results:

• FreqMamba-Net demonstrated superior segmentation accuracy and boundary consistency compared to existing models.
• The framework showed strong correlation with clinical severity proxies.
• Achieved robust generalization across diverse colonoscopy imaging scenarios.

Conclusions:

• FreqMamba-Net effectively identifies diffuse inflammatory lesions difficult for conventional methods.
• This framework offers a scalable foundation for quantitative, objective IBD assessment tools in clinical practice and trials.