信息工程与人工智能学院知识库

结直肠癌是一种常见且致死率较高的全球性癌症。及早进行息肉切除手术可以显著降低癌变的风险，因此定期进行结直肠检查对于预防结直肠癌非常重要。目前，结肠镜检查是最常用的结直肠息肉检测技术。然而，由于息肉的形态各异、大小不一，并且与粘膜的边界模糊，准确地分割息肉对于计算机来说仍然是一个巨大的挑战。随着深度学习的发展，许多息肉分割方法应运而生，在医学实践中极大地帮助医生寻找和识别结直肠息肉。然而，大多数研究使用通用的传统模型，很少有研究考虑到息肉分割和伪装物体检测之间的联系。因此，本文分别从全监督和半监督的角度出发来解决问题。具体而言，本文的主要研究内容包括以下两个方面：
（1）基于卷积反向注意力模块的息肉分割网络：考虑到息肉的边界与结直肠壁之间的模糊不清，本文借鉴了伪装物体检测的思路中的反向注意力来解决息肉分割的问题。反向注意力可以预测出伪装目标的边缘特征，以往的研究未能考虑到卷积注意力模块（Convolutional Block Attention Module，CBAM）的定位性能，因此本文将其引入到卷积反向注意力中。首先，将息肉图像输入特征提取网络进行特征提取，并利用特征金字塔（Feature Pyramid，FP）模块增强特征。然后，通过聚合解码器（Aggregation Decoder，AD）模块获得粗略的分割结果。最后，利用卷积反向注意力（Convolution Reverse Attention，CRA）模块对粗略的分割结果进行细化。相比几种经典的息肉分割网络，本文提出的方法能够提高分割的准确性。
（2）基于mean teacher的半监督息肉分割网络：鉴于标注数据对于息肉分割任务来说既耗时又昂贵，而无标注的息肉分割数据相对丰富，本文引入了半监督的思路，充分利用无标注数据。具体而言，一方面，本文利用基于数据变换的正则化方法提高网络的鲁棒性。另一方面，引入不确定性感知机制，使得学生模型能够学习到更可靠的信息。在常见的公开数据集上，本文提出的方法在半监督网络性能方面优于经典的半监督网络。

Colorectal cancer is a common and deadly cancer worldwide. Early polyp removal surgery can significantly reduce the risk of cancer, so regular colon exams are important to prevent colorectal cancer. Currently, colonoscopy is the most commonly used technique for detecting colorectal polyps. However, because polyps vary in shape, size, and have blurred boundaries with mucous membranes, accurately partitioning polyps remains a huge challenge for computers. With the development of deep learning, many polyp segmentation methods have emerged, which greatly help doctors find and identify colorectal polyps in medical practice. However, most studies have used generic conventional models, and few have considered the link between polyp segmentation and camouflaged object detection. Therefore, this paper starts from the Angle of full supervision and semi-supervision to solve the problem. Specifically, the main research contributions of this paper are as follows:
(1) Polyp segmentation network based on convolutional reverse attention module: Considering the ambiguity between the polyp boundary and the colorectal wall, this paper uses the reverse attention in the idea of camouflage object detection to solve the problem of polyp segmentation. While reverse Attention can predict the edge features of camouflaged objects, previous studies fail to take into account the positioning performance of the Convolutional Block Attention Module (CBAM), so this paper introduces it into the convolutional block attention module. Firstly, the polyp image was input into the Feature extraction network for feature extraction, and the Feature Pyramid (FP) module was used to enhance the features. Then, rough segmentation results are obtained through the Aggregation Decoder (AD) module. Finally, we use the Convolution Reverse Attention (CRA) module to refine the rough segmentation results. Compared with several polypus segmentation networks, the proposed method can improve the accuracy of segmentation. 
(2) Semi-supervised polyp segmentation network based on mean teacher: Considering that labeling data for polyp segmentation tasks is time-consuming and expensive, while unlabeled polyp segmentation data is relatively abundant, this paper introduces a semi-supervised approach to fully utilize unlabeled data. Specifically, on one hand, this paper improves the network's robustness using regularization methods based on data transformations. On the other hand, an uncertainty perception mechanism is introduced, enabling the student model to learn more reliable information. On commonly used public datasets, the proposed method in this paper outperforms classical semi-supervised networks in terms of semi-supervised network performance.