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

Colorectal cancer (CRC) is a leading cause of cancer-related mortality globally. The early detection and resection of adenomatous polyps via colonoscopy are critical for improving survival rates. However, missed diagnosis rates remain suboptimal, primarily due to the subtle appearance of flat polyps and the variable complexity of the intestinal environment. While deep learning has shown promise in automated polyp segmentation, existing models often face a trade-off between high segmentation accuracy and the real-time inference speeds required for clinical deployment.

To address this challenge, we propose PolyMamba-Net, a novel hybrid architecture that synergizes the efficient long-range dependency modeling of State Space Models (Mamba) with the local feature extraction capabilities of Convolutional Neural Networks (CNNs). Specifically, we introduce a dual-branch encoder designed to capture both global context and fine-grained textures, coupled with a Boundary-Aware Module (BAM) to explicitly refine polyp margins. A composite loss function targeting structural, pixel-level, and boundary consistency was employed for training.

Extensive experiments were conducted on two public benchmarks: Kvasir-SEG and CVC-ClinicDB. PolyMamba-Net achieved a Dice Coefficient of 0.942 and 0.935, respectively, significantly outperforming state-of-the-art methods such as Swin-UNet and PraNet. Furthermore, our model surpasses recent 2024-2025 approaches including ADPNet and AAPCNet across all evaluation metrics, with all improvements confirmed as statistically significant (p < 0.05, Wilcoxon signed-rank test). Notably, our model operates at 115 FPS on a single NVIDIA RTX 3090 with only 25.3M parameters and 12.8 GFLOPs, demonstrating significantly higher efficiency than transformer-based counterparts. Cross-dataset evaluations on three unseen benchmarks further validate the model's generalizability.

PolyMamba-Net demonstrates superior segmentation precision combined with real-time processing capability. It offers a clinically feasible solution for assisting endoscopists in minimizing missed detection rates during routine colonoscopies.