• 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • br Introduction br Cancer is a critical


    1. Introduction
    Cancer is a critical public health problem worldwide. Among the cancer types, breast cancer incidence rates are second highest for women, excluding lung cancer. In addition, the mortality of breast cancer is very high when compared to other types of cancer [1]. Even with the rapid advances in medicine, the analysis of histopathological images remains the most widely used method for breast cancer diag-nosis [2]. In all the histopathological image analysis tasks, the most important is the classification task. Because the automatic and precise classification of high-resolution histopathological images is the cor-nerstone and bottleneck of other in-depth studies, such as nuclei loca-lization, mitosis detection and gland segmentation.
    Currently, histopathological imaging in clinical practice is mainly based on the manual qualitative analysis of pathologists. However, at least three problems arise from this analysis method. First, there is a shortage of pathologists in the world, especially in less developed areas
    and small hospitals. This resource shortage and unbalanced distribution is an urgent problem to be solved. Second, whether the histopatholo-gical diagnosis is correct or not correct completely depends on the pathologist's profound professional knowledge and long-term accumu-lated diagnostic experience. This pathologist subjectivity has led to a proliferation of diagnostic inconsistencies. Third, the complexity of the histopathological images makes pathologists prone to fatigue and in-attention. Facing these problems, it is urgent to develop automatic and precise histopathological image analytical methods, especially classifi-cation methods, to alleviate these problems.
    Recently, deep learning methods have made considerable progress and achieved remarkable performance in the field of computer vision and image processing, which has inspired many scholars to apply this technique to histopathological image classification [3]. Convolutional neural networks (CNNs) are the most widely used type of deep learning network, and they perform equally well on image classification and image feature extraction [4]. These results have laid the foundation for
    Corresponding authors.
    the application of the CNN in histopathological image classification. However, in Nicotinamide riboside chloride to natural images, histopathological images are characterized by high resolution. Limited by the memory of the gra-phics processing unit (GPU), it is impossible to feed high-resolution images as a whole into the CNN for classification. Additionally, it is impractical to simply resize a high-resolution image into a low-resolu-tion image because considerable useful image information will be lost. This is especially true in medical imaging, where data volumes are al-ready small.
    The mainstream method of histopathological image classification divides a whole image into smaller patches, then uses a CNN to classify each patch, and finally integrates the classification results of these patches, such as majority voting, to determine the classification results. A CNN is also used to extract the feature representation vector of each patch, and then the traditional machine learning classification algo-rithm, such as support vector machine (SVM), is used to make the classification result of the whole histopathological image [5].
    However, the current mainstream method faces three challenges. First, the high-resolution characteristics of histopathological images have not been fully utilized to improve the classification accuracy but have caused great negative effects. The main reason is that the current best patch-based method does not adequately integrate these patches to make the classification result of the whole histopathological image. Specifically, these methods integrate only the short-distance de-pendency between patches but ignore the long-distance spatial de-pendency, which is very helpful for the context understanding of the whole image. Second, the feature representation of the pathological image patch is not sufficiently richer. Thus, a large amount of in-formation is lost before image-wise fusion, making the fusion in-sufficient. Finally, there are considerable challenges in terms of data-sets. Many very important advances in computer vision fields have benefited from an open research environment enabled by publicly available datasets for benchmarking, such as ImageNet, for object re-cognition in natural images. Medical imaging researchers have even-tually started to follow this lead with the release of well-annotated datasets such as the BreaKHis dataset [6] and Bioimaging2015 dataset [7]. However, these datasets are still relatively small. In particular, the diversity of the dataset is not guaranteed.