PolyMamba-Net: a lightweight and boundary-aware network for real-time polyp segmentation in colonoscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer (CRC) is a major global health concern, with early detection crucial for survival.
• Colonoscopy is key for identifying and removing precancerous polyps, but missed diagnoses, especially of flat polyps, remain an issue.
• Current deep learning models for polyp segmentation struggle to balance accuracy with the real-time processing needed for clinical use.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for automated polyp segmentation in colonoscopy images.
• To address the limitations of existing models in handling subtle polyp appearances and achieving real-time inference speeds.

Main Methods:

• Proposed PolyMamba-Net, a hybrid architecture combining State Space Models (Mamba) for long-range dependencies and Convolutional Neural Networks (CNNs) for local features.
• Introduced a dual-branch encoder for global and local feature extraction and a Boundary-Aware Module (BAM) for precise polyp margin delineation.
• Utilized a composite loss function for structural, pixel-level, and boundary consistency during training.

Main Results:

• PolyMamba-Net achieved high Dice Coefficients (0.942 on Kvasir-SEG, 0.935 on CVC-ClinicDB), outperforming state-of-the-art methods including recent 2024-2025 approaches.
• Demonstrated statistically significant improvements (p < 0.05) over competitors across all metrics.
• Achieved real-time performance (115 FPS) with a compact model size (25.3M parameters, 12.8 GFLOPs), surpassing transformer-based models in efficiency.

Conclusions:

• PolyMamba-Net offers a clinically viable solution for enhancing polyp detection during colonoscopies.
• The model's high segmentation accuracy and real-time processing capabilities can significantly assist endoscopists in reducing missed polyp diagnoses.