Backgrounds <p>Colonoscopy plays a crucial role in preventing the malignant transformation of colorectal polyps, with early diagnosis and detection of colorectal cancer being effective approaches to reducing incidence and mortality rates among patients. With the rise of neural networks, research on computer-aided detection of colorectal polyps has garnered increasing attention. However, existing computer-aided diagnostic systems are constrained by insufficient training sample sizes, making it difficult to train high-performance systems.</p> Methods <p>This paper proposes the Y-Polyp model for colorectal polyp detection. By employing a parallel strategy with multidimensional attention mechanisms, this model enables convolutional kernels to learn more flexible attention across four spatial dimensions, thus fully capturing target features from limited data samples. Additionally, the model filters conflicting information in spatial domains, suppressing inconsistent features and addressing inconsistencies among features at different scales. To validate the effectiveness of the Y-Polyp model, extensive experiments are conducted and evaluated using a Kvasir-SEG dataset.</p> Results <p>The experimental results show that the Y-Polyp model significantly improves the accuracy in the detection of colorectal polyps. Specifically, the overall precision of the polyp detection model has increased by 3%, while the recall rate and average precision have increased by 2.7%, 2% and 2.7%, respectively.</p> Conclusions <p>This method exhibits strong generalization ability and robustness, adapting to various colonic structures and changes in target appearance, and the proposed colorectal polyp detection approach thus possesses wide applicability and reliability.</p>