Attention-Guided U-Net With Grad-CAM for Explainable Polyp Segmentation in Colonoscopy Images

Colorectal cancer prevention relies on early detection and accurate segmentation of polyps in colonoscopy images, yet challenges such as diverse polyp morphologies, varying imaging conditions, and lack of model interpretability hinder effective diagnosis. This study proposes an enhanced U-Net architecture integrated with attention gates to prioritize salient polyp regions and Gradient-weighted Class Activation Mapping (Grad-CAM) to provide visual explanations of decision-critical areas, ensuring both precision and clinical transparency. Trained on the Kvasir-Polyp dataset with advanced data augmentation and refined annotation strategies, the model leverages attention mechanisms and explainable AI to address variability in lighting, resolution, and polyp characteristics. The framework demonstrates robust segmentation performance through rigorous 5-fold cross-validation, with external validation on the CVC-ClinicDB dataset confirming its generalizability across diverse clinical imaging environments. Comparative analysis with state-of-the-art models, including Swin-Unet, TransUNet, and UNETR++, reveals a superior balance of accuracy, computational efficiency, and inference speed, making it ideal for real-time clinical workflows. Grad-CAM heatmaps align closely with expert-annotated ground truth, reinforcing clinical trust and interpretability. The proposed framework demonstrates strong potential for integration into real-time clinical workflows, combining high segmentation performance with interpretability to support diagnostic decision-making.