基于卷积神经网络和胶囊神经网络的宫颈病变图像分类方法研究

为了更好地治疗宫颈癌,准确确定患者的宫颈类型是至关重要的。因此,用于检测和划分宫颈类型的自动化方法在该领域中具有重要的医学应用。虽然深度卷积神经网络和传统的机器学习方法在宫颈病变图像分类方面已经取得了良好的效果,但它们无法充分利用图像和图像标签的某些关键特征之间的长期依赖关系。为了解决这个问题,文章引入了胶囊网络(CapsNet),将CNN和CapsNet结合起来,以提出CNN-CapsNet框架,该框架可以加深对图像内容的理解,学习图像的结构化特征,并开展医学图像分析中大数据的端到端训练。特别是,文章应用迁移学习方法将在ImageNet数据集上预先训练的权重参数传输到CNN部分,并采用自定义损失函数,以便网络能够更快地训练和收敛,并具有更准确的权重参数。实验结果表明,与ResNet和InceptionV3等其他CNN模型相比,文章提出的网络模型在宫颈病变图像分类方面更加准确、有效。

Research on image classification of cervical lesions based on convolution neural network and capsule neural network

In order to better treat cervical cancer, it is very important to accurately determine the cervical type of patients. Therefore, the automatic methods for detecting and classifying cervical types have important medical applications in this field. Although deep convolution neural network(CNN) and traditional machine learning methods have achieved good results in cervical lesions image classification, they can not make full use of the long-term dependence between some key features of image and image tag. To solve this problem, the CapsNet is introduced. Specifically, this paper combines CNN and CapsNet to propose the CNN-CapsNet framework, which can deepen the understanding of image content, learn the structural features of images, and carry out end-to-end training of big data in medical image analysis. In particular, this paper uses the migration learning method to transfer the weight parameters trained in advance on the Imagenet dataset to CNN, and use the custom loss function, so that our network can train and converge faster and have more accurate weight parameters. The experimental results show that compared with other CNN models such as ResNet and inception V3, our network model is more accurate and effective in cervical lesions image classification.