目的 在医疗保健领域，复杂多变的背景分布、息肉形态与尺寸的显著差异，以及边界定义的模糊性为实现息肉的精确分割造成诸多挑战。为应对上述难题，本文创新性地提出了一种针对结肠镜检查过程中息肉分割问题的深度学习模型，即息肉边界线索深度融合网络（Polyp Boundary Cues Deep Fusion Network，PBCDF-Net）。方法 本文所提出的PBCDF-Net网络使用Res2Net-50作为骨干网络，并设计了一种边界线索挖掘模块（Boundary clue mining module，BCMM），旨在合并从骨干网络派生的多级特征，以提取隐藏的边界细节。此外，本文使用前景目标增强模块（Foreground target enhancement module，FTEM）来增强网络对前景目标的关注。最后，在解码阶段设计了一种深度特征融合模块（Deep feature fusion module，DFFM）来整合提取的边界信息和前景目标信息。结果 在本研究中，我们以五个公共数据集（Kvasir、ETIS、CVC-ColonDB、CVC-ClinicDB和CVC-300）作为测试基准，全面评估了所提出的PBCDF-Net模型在结直肠息肉分割任务上的性能，并在最新的数据集PolypGen上进行了one-in-out的交叉实验。具体的，在CVC-ClinicDB数据集上，PBCDF-Net与CCBANet相比，在五项评价指标上分别提升了6.6%、7.4%、3.4%、7%和4.9%。在Kvasir和CVC-300数据集上，与近几年方法相比，PBCDF-Net在所有评估指标上平均提升了4.5%、6.2%、2.5%、6.3%和2.9%。此外，PolypGen数据集上的交叉实验结果表明，与PraNet相比，PBCDF-Net在mDice和mIOU上分别提高了4.6%和4.9%，并且在个别指标上优于最先进的方法。结论 本文提出的息肉图像分割网络（PBCDF-Net）成功克服了传统算法在面对息肉边界不确定性及形态多样性时的局限性，在广泛的公开数据集验证中展现出了卓越的性能表现，特别是在处理边界模糊且形态多变的复杂息肉案例时，其分割精度与鲁棒性远超同类方法。

Objective Colorectal cancer, as a high-incidence and extremely harmful disease, poses a severe challenge to human health. According to statistics, about 95% of cases originate from the progressive development of colon polyps, highlighting the importance of early identification and monitoring of polyps, which is crucial to significantly reducing the incidence of colorectal cancer. However, traditional manual diagnosis methods have a high omission rate, which limits the effectiveness of early intervention. In this context, the introduction of deep learning technology provides a new way to solve this problem. By finely analyzing the characteristics of lesions, such as the precise location and morphological structure of polyps, deep learning models can significantly enhance doctors' screening efficiency and accuracy, bringing innovation to the prevention and treatment of colorectal cancer. In recent years, with the rapid progress of deep learning technology, its application in medical image analysis and other fields has achieved breakthrough progress, especially models such as convolutional neural network (convolutional neural network,CNN) and visual transformer (visual transformer,ViT), which have been widely used in medical task processing. It has demonstrated excellent performance and further promoted the clinical popularization trend of computer-aided diagnosis technology. In view of the complex characteristics of colorectal polyp images such as significant morphological heterogeneity and blurred edge definitions, this study innovatively designed a polyp boundary clue deep fusion network (polyp boundary clue deep fusion network,PBCDF-Net), focusing on improving the segmentation accuracy of polyp images. This network achieves accurate capture and segmentation of polyp boundaries by integrating multi-layer features. It has been verified by multiple types of data sets and has demonstrated excellent performance. It not only deepens the understanding of the pathological characteristics of polyps, but also provides a powerful tool for clinical practice. A tool with important practical value and forward-looking perspective. Methods The proposed PBCDF-Net network uses Res2Net-50 as the backbone network. By extracting different receptive field features, it enhances the multi-scale representation ability of the target object, so that the network has strong feature extraction capabilities and model expression capabilities. Aiming at the fuzziness and uncertainty of polyp boundaries, this paper specially designed a boundary clue mining module (boundary clue mining module,BCMM) to extract effective boundary clues from low-level feature layers containing rich texture detail information by embedding specific operators in the module. , and combined with advanced semantic feature layers to accurately locate polyp locations, thereby obtaining more accurate and effective boundary information. Afterwards, the mined boundary clues are fused with different levels of semantic feature layers to achieve higher-precision polyp fuzzy boundary segmentation by making up for the lack of boundary detail information in the semantic feature layer, thereby further improving the model segmentation performance. In view of the significant morphological differences and structural complexity of polyps themselves, we designed a foreground target enhancement module (foreground target enhancement module,FTEM) that can enhance target features for difficult-to-find small target polyps and polyps with complex structures to promote the network The ability to identify and perceive polyps with different tissue structures. In order to efficiently combine boundary detail features and target enhancement features, in the decoding stage, we designed a deep feature fusion module (deep feature fusion module,DFFM), which performs a preliminary fusion of the two features and then performs hierarchical cross-over of the preliminary fused features. Fusion, ultimately achieving deep fusion of features, DFFM is performed in the form of cascade transfer fusion, achieving reliable upper and lower feature correlation. In terms of data set processing, this article uses experimental data configurations of various mainstream networks such as PraNet. In detail, our training data has a total of 1450 polyp images, which are composed of 900 images from the Kvasir dataset and 550 images from the CVC-ClinicDB dataset. Then, we combine the remaining data of the Kvasir data set and CVC-ClinicDB data set with the ETIS data set, CVC-ColonDB data set, and CVC-300 data set to form test data. In terms of model performance evaluation, this article uses five indicators to evaluate the model, including average Dice, average IOU, structure measure, weighted F measure, and enhanced alignment measure. Results In this study, we comprehensively evaluated the performance of the proposed PBCDF-Net model on the colorectal polyp segmentation task using five public datasets (Kvasir, ETIS, CVC-ColonDB, CVC-ClinicDB, and CVC-300) as benchmarks for testing, and additionally conducted one-in-out crossover experiments on the latest datasets PolypGen. These datasets cover a wide range of morphological features of polyps, ensuring the breadth and validity of the evaluation. In order to objectively evaluate the progress of PBCDF-Net, we conducted a systematic comparative analysis of its performance using nine advanced segmentation methods, including UNet, UNet++, SFA, PraNet, ACSNet, CCBANet , DCRNet, ECTransNet and CIFGNet. The experimental results clearly demonstrate the excellent segmentation capabilities of PBCDF-Net under various data sets. In particular, on the CVC-ClinicDB dataset, compared to CCBANet, PBCDF-Net shows an increase in the important metrics mDice, mIOU , structural metrics, weighted F-measure, and augmented alignment metrics by 6.6%, 7.4%, 3.4%, 7%, and 4.9%, respectively. Similarly, on the Kvasir and CVC-300 datasets, PBCDF-Net exhibits an average of 4.5%, 6.2%, 2.5%, 6.3%, and 2.9% improvement in all evaluation metrics compared to the methods of recent years. In addition, cross-experiment results on the PolypGen dataset show that PBCDF-Net improves 4.6% and 4.9% on mDice and mIOU, respectively, compared to PraNet, and outperforms the state-of-the-art methods on several metrics. These improvements not only reflect PBCDF-Net's strong ability to maintain segmentation structural integrity and detail fidelity, but also demonstrate its significant advantages in improving the quality of segmentation output. Conclusion The constructed PBCDF-Net model demonstrated excellent performance in colorectal polyp segmentation tasks. Through careful subjective evaluation, the network's consistent performance on multiple datasets was confirmed, highlighting its strong adaptability to polyp size diversity and edge fuzziness, as well as its high performance in accurately defining polyp contours. In addition, the ablation experimental analysis of the three core components in the network design (boundary information mining module, foreground region enhancement module, and deep feature integration module) clearly confirmed that these components played a decisive role in enhancing the model's segmentation accuracy, which effectively improved the overall performance of the algorithm.