Cascaded attention-driven dynamic multi-scale feature fusion network for gastrointestinal polyp segmentation

Intestinal endoscopic images are frequently degraded by uneven illumination, motion blur, and specular reflections, which tend to cause diffuse lesion boundaries and obscure micro-lesions. Moreover, conventional multi-scale feature fusion methods struggle to accommodate the drastic scale variations of lesions induced by fluctuating shooting distances. To address these challenges, this paper proposes a Cascaded Attention-Driven Dynamic Multi-Scale Feature Fusion Network(CADM-Net). The network first constructs a Multi-Kernel Residual Block to perform parallel multi-scale feature extraction, thereby effectively preserving the fine-grained details of micro-lesions. Subsequently, a Dynamic Multi-Scale Attention mechanism is introduced, which leverages learnable context aggregation and serial dual-attention calibration to adaptively perceive and respond to severe lesion scale variations. Finally, a dedicated Feature Fusion Module is designed to alleviate cross-level semantic misalignment and further refine diffuse lesion boundaries. These modules operate synergistically to significantly enhance the network’s robustness against endoscopic imaging degradation and ensure precise segmentation of lesion regions across varying scales. Extensive experimental results on four public polyp benchmark datasets demonstrate that CADM-Net consistently surpasses existing state-of-the-art methods, achieving superior segmentation accuracy and boundary delineation under diverse imaging conditions.