There are many different types of nuclei in a tumor tissue. We can identify the specific nuclei and their distribution in the tissue to reflect the current cancer state of histopathological images. However, due to the existence of cellular heterogeneity, the recognition of nuclei in histopathological images has always been a problem of computer vision. In the paper, we use the transfer learning of Deep Convolutional Neural Network to classify nuclei, and found that adjusting the size of the nuclear image to a certain size can improve the accuracy of the nuclei classification model, while not significantly reducing the nuclei classification efficiency. Through further research, it was confirmed that the environment around the nuclei can bring great help to the model classification. Based on the principle, we design a feature fusion model. We extract features from nuclei image different sizes by CNN, fuse the features, and then use fully connected layer to classify the features. Experiments have proved that the feature fusion model has a considerable improvement in accuracy compared to the normal classification model.

中文翻译：

---

基于深度学习的癌核图像数据分类与特征融合方法研究

肿瘤组织中有许多不同类型的细胞核。我们可以识别特定的细胞核及其在组织中的分布，以反映组织病理学图像的当前癌症状态。然而，由于细胞异质性的存在，组织病理学图像中细胞核的识别一直是计算机视觉的难题。在论文中，我们使用深度卷积神经网络的迁移学习对细胞核进行分类，发现将细胞核图像的大小调整到一定大小可以提高细胞核分类模型的准确率，同时不会显着降低细胞核分类效率. 通过进一步的研究，证实了原子核周围的环境对模型分类有很大帮助。基于该原理，我们设计了一个特征融合模型。我们通过CNN从不同大小的核图像中提取特征，融合特征，然后使用全连接层对特征进行分类。实验证明，特征融合模型与普通分类模型相比，在准确率上有相当大的提升。