融合几何先验与双域特征提纯的结直肠息肉分割网络OA

To address the challenges of tiny lesion segmentation,specular reflection interference,and inac-curate boundary localization during colorectal polyp segmentation,this paper proposes a Geometric Prior and Dual-Domain Purification Network(GPDP-Net).First,an Optimized Edge Module(OEM)is de-signed to extract explicit geometric priors,and an Improved Coordinate Attention Mechanism is intro-duced to enhance spatial position awareness.Second,a Parallel Dual-Domain Enhancement Hub(PD-DEH)is designed,where the evolutionary constraint based on the Physics-Informed Neural Network(PINN)is employed to suppress incoherent optical noise,while the Fast Fourier Transform Branch is uti-lized to recover high-frequency information in the frequency domain.Finally,a Sub-pixel Restoration and Iterative Reverse-Attention Decoder(SR-IRAD)is incorporated into the cascaded decoder to achieve high-fidelity restoration of lesion details.Experimental results on four benchmark datasets demonstrate that GPDP-Net achieves strong performance on various evaluation metrics.Specifically,on the ETIS-Larib-PolypDB dataset containing numerous tiny lesions,the proposed network achieves an mIoU of 0.701,which is 1.4 percentage points higher than the second-best method.These results demonstrate that GPDP-Net can effectively improve the segmentation accuracy of colorectal polyps with complex structures and en-hance the robustness and generalization capability of the model in challenging clinical scenarios.