目的 结肠镜技术在结肠癌的早期检测中发挥着重要作用,但需要操作员的专业技能和主观判断,因而存在较大局限性。现有结肠癌图像分割方法大多数采用额外层和显式扩展网络结构,导致模型效率低下。此外,由于息肉与其周围粘膜之间的边界不清晰,现有模型对于息肉边界的分割效果并不理想。方法 本文提出一种端到端的自知识蒸馏框架,专门用于结肠癌图像分割。该框架将边界分割网络和息肉分割网络整合到统一知识蒸馏框架中,以相互增强两个网络的性能。采用专注于边界分割的模型作为教师网络,将息肉分割模型作为学生网络,两者之间共享一个特征提取模块,以促进更有效的知识传递。设计了一种反向特征融合结构,通过上采样以及矩阵乘法聚合深层特征,并利用反向浅层特征作为辅助信息,从而获得分割掩模的全局映射。结果 通过在CVC-ClinicDB、CVC-ColonDB、Kvasir以及HAM10000等四个数据集上开展实验,实验结果表明提出的模型在当前最先进的分割方法中具有最佳效果,DSC和mIoU指标比最好模型分别提升了0.45%和0.68%。 结论 提出的模型适用于各种大小和形状的息肉分割,实现准确的边界提取,并且具有推广到其他医学图像分割任务的潜力。

Objective Colorectal cancer remains a formidable global health challenge, underscoring the pressing need for early detection strategies to improve treatment outcomes. Among these strategies, colonoscopy stands out as a primary diagnostic tool, relying on the visual acumen of medical professionals to identify potentially cancerous abnormalities, such as polyps, within the colon and rectum. However, the effectiveness of colonoscopy is heavily contingent upon the skill and experience of the operator, leading to variability and limitations in detection rates across different practitioners and settings. In response to these challenges, the integration of artificial intelligence (AI) and computer vision techniques has garnered increasing attention as a means to augment the accuracy and efficiency of colorectal cancer screening. Various algorithms have been developed to automatically segment colorectal images, with the overarching goal of precisely delineating polyps from the surrounding tissue. Despite advancements in this domain, many existing models confront inherent inefficiencies and limited effectiveness stemming from their intricate architectures and dependence on manual feature engineering. Method This paper proposes a novel approach termed Boundary self-Knowledge Distillation (BA-KD), which aims to achieve precise polyp segmentation in a comprehensive, end-to-end fashion. Unlike conventional methods, BA-KD seamlessly integrates boundary and polyp segmentation networks into a unified framework, facilitating effective knowledge transfer between the two domains. Notably, the concept of Boundary self-Knowledge Distillation represents a pioneering contribution in this field, aiming to harness the synergistic benefits of both boundary and polyp information for enhanced segmentation accuracy. The BA-KD framework comprises two interconnected branches: a boundary segmentation network serving as the teacher branch and a polyp segmentation network acting as the student branch. To address the inherent challenges associated with delineating polyp boundaries, a boundary detection operator is introduced to automatically generate boundary masks, which are subsequently leveraged in training both branches. This approach not only enhances the segmentation performance of the student branch but also enriches the knowledge base of the teacher branch, thereby fostering mutual learning and refinement. A key distinguishing feature of BA-KD is the shared image feature extractors between the student and teacher branches, facilitating robust knowledge transfer across both domains. Moreover, to facilitate effective fusion of feature information at various hierarchical levels, two innovative structures are proposed: Reverse Multi-Level Feature Fusion (RMLF) and Reverse Feature Fusion (RFM). RMLF enables the integration of high-level features to generate a comprehensive global feature map, while RFM synergistically combines reverse shallow features with high-level features aggregated through RMLF to produce the final segmentation mask. Result The paper conducts comprehensive experimental validation of BA-KD"s results against seven state-of-the-art methods across four different datasets. These datasets include CVC-ClinicDB, CVC-ColonDB, Kvasir, and HAM10000. The comparative algorithms consist of U-Net, Double-UNet, UNet++, TransFuse, PraNet, DuAT, and RaBit. These algorithms serve as benchmarks in polyp segmentation and general medical image segmentation domains. Results on CVC-ClinicDB: BA-KD demonstrates exceptional performance in terms of mSpe and mDSC, achieving 0.997 and 0.9555, respectively. BA-KD outperforms all competitors in terms of mDSC and mIoU, with improvements of 0.45% and 0.68% compared to RaBit, respectively. Results on CVC-ColonDB: On this dataset, BA-KD surpasses all other methods across all evaluation metrics. Compared to the optimal performance achieved by TransFuse, our BA-KD achieves improvements of 2.20% in mIoU and 1.51% in mDSC. Results on Kvasir: For the Kvasir dataset, BA-KD achieves an mIoU of 0.889 and an mDSC of 0.937, surpassing the best-performing RaBit by approximately 1.08% and 1.14%, respectively. Furthermore, the paper further evaluates BA-KD"s generalization ability on other medical segmentation tasks (HAM10000, which includes dermoscopic images from different populations). Compared to existing medical segmentation baselines, BA-KD excels in all metrics on HAM10000. BA-KD achieves significant scores in mDSC (0.9562) and mIoU (0.9223), surpassing the best-performing DoubleU-Net by 1.45% and 2.25%, respectively. The experimental results clearly demonstrate that BA-KD exhibits superior performance compared to existing state-of-the-art segmentation methods, with significant improvements in Dice similarity coefficient (DSC) and mean Intersection over Union (mIoU) metrics. Conclusion In summary, the BA-KD framework represents a promising advancement in the field of colorectal cancer screening, offering a robust and accurate solution for the segmentation of polyps across varying sizes and shapes. Beyond its immediate application in colorectal cancer diagnosis, the adaptability and versatility of BA-KD suggest its potential utility in other medical image segmentation tasks, thereby contributing to the broader goal of advancing computer-aided diagnosis systems and ultimately improving patient outcomes in the realm of medical imaging.