PolypSegNet: An Attention-Guided Multi-Scale Feature Fusion Network for Accurate Colorectal Polyp Segmentation

Colorectal cancer (CRC) is a leading cause of
cancer-related mortality worldwide, making early colorectal
polyp detection essential for improving patient outcomes. However, accurate polyp segmentation remains challenging due to
variations in polyp size, shape, and texture, along with low
contrast, blurred boundaries, and imaging artifacts in endoscopic
images. Although recent CNN-, Transformer-, and foundation
model-based approaches have improved segmentation accuracy,
many remain limited by high computational cost and architectural complexity. To address these issues, this study proposes
PolypSegNet, an efficient encoder–decoder architecture integrating a pretrained ResNet-50 encoder, Multi-Scale Feature Fusion
(MSFF), and Convolutional Block Attention Modules (CBAM).
The framework enhances multi-scale contextual learning and
attention-guided feature refinement while maintaining computational efficiency. Experiments on the Kvasir-SEG dataset achieved
a Dice score of 0.9164 and an IoU score of 0.8655, demonstrating
competitive performance for colorectal polyp segmentation.