LiteSegNet: Fusing Lightweight Mechanisms for Efficient Multi-Scale Polyp Segmentation

Accurate polyp segmentation during colonoscopy is crucial for colorectal cancer prevention, yet existing methods struggle with polyp morphological variability and computational efficiency. To address this, we propose LiteSegNet, a lightweight network for efficient polyp segmentation. The model integrates key innovations including an Optimized ASPP Module employing multi-dilation-rate convolutions and global pooling to capture multi-scale contextual features; an Adaptive Multi-Scale Feature Fusion (ASFF) Module that dynamically weights and fuses low-, middle-, and high-level features to enhance detail preservation; and a Shifted MLP Module combining spatial shifting with depth-wise separable convolutions (DWConv) to strengthen spatial dependencies while minimizing computational cost. Evaluated on the Kvasir-SEG and CVC-ClinicDB datasets, LiteSegNet achieves state-of-the-art performance (Dice: 85.69%/88.94%; mIoU: 78.63%/82.80%) with only 3.3M parameters and 4.30 GFLOPs, significantly outperforming models like U-Net++ and UNext in both accuracy and efficiency.