A cognitive dual-attention network with feature specificity for automated CRC polyp detection

Colorectal cancer is the foremost cause of cancer-related deaths worldwide. Thus, early diagnosis of precancerous polyps is crucial for more effective treatment outcomes. To address the persistent issues in colorectal cancer diagnosis, this study proposes a novel classification model, E-D²AN (Efficiently Dilated Dual Attention Network), capable of accurately distinguishing colorectal abnormalities from normal colonoscopy images. The proposed model features dual attention mechanisms of EDCAM (Efficiently Dilated Channel Attention Mechanism) and ESAM (Efficient Spatial Attention Mechanism) strategies, utilizing an extended ResNet50 as the backbone. This network prioritizes efficient attention mechanisms, enriching feature extraction through an expanded receptive field and significantly driving attention-focused learning to achieve improved precision in the classification model. This combination improves the model’s ability to locate and focus on crucial regions in images, resulting in higher diagnostic precision. Dropblock regularization is also used strategically to reduce overfitting and improve generalization to unseen images. The proposed E-D²AN model excels across all three benchmark datasets, demonstrating superior performance over existing approaches. Prominently, it achieves an accuracy of 83.69 % on PolypsSet, 99.74 % on CKC, and 98.75 % on Kvasir-2 datasets. These findings reveal the model’s ability to improve the accuracy and reliability of early polyp detection.