The systematic study of nuclei patterns in pathology images is very important for fully charactering the grade of cancerous tissues. Nowadays, with the great advance of deep neural networks (i.e., DNN), intense interest in adopting DNN to distinguish different types of pathology nuclei is widely spread. However, most of the existing methods need to annotate lots of nuclei images in the training stage, and this is not always an option for the labelling cost are high. To address this problem, we propose a novel approach called DAPC (i.e., deep active learning with pairwise constraints) to actively select the most valuable nuclei for annotation. Specifically, we firstly design a novel pairwise-constraint regularized deep convolutional neural network (i.e., CNN) that can simultaneously preserve the distribution of different subjects and optimize the objective criterion of conventional CNN. Then, through the properly designed CNN, we query the most informative nuclei in the unlabelled dataset for human annotation, and the parameters of the designed CNN is subsequently updated by incorporating the newly annotated samples to enhance the CNNs performance incrementally. We evaluate our method on a public available pathology colon dataset, the experimental results show that the proposed method could achieves to a weighted F1-score of 79.2% by only annotating 60% nuclei in the training set, which is better than the comparing methods.

INTRODUCTION

It is widely recognized that the characteristics (i.e., size ,shape and texture) of nuclei are important factors in diagnosing the grade of cancerous tissues [1-2], and thus one important task in the research of pathology is to distinguish different types of nuclei at cellular level [3]. Recently, with the breakthrough of microscopic imaging technology, scientists are collecting large volumes of hematoxylin-eosin staining images to analysis their covered nucleus patterns[4][5]. Hence, finding an automatic computational way to complete the nuclei classification task has been becoming a new focus in pathology. From the perspective of machine learning, the task of distinguishing different types of nuclei can be treated as a classification problem, and most of the existing methods have solved it by adopting a two-step framework, where they firstly endeavor to figure out a proper feature representation way for encoding the image data, which then will be fed into an appropriate classifier for category decision. For instance, Cosatto et al [4] have used the shape and texture feature for nuclear image description and a AdaBoost classifier for nuclei pleomorphism grading, Yuan et al [6] have adopted SVM classifier to classify nucleus into cancer, lymphocyte or stromal based on the morphological features Nowadays, with the great advance of DNN [7], deep learning approaches have been shown to produce much more satisfactory results than the traditional methods on the task of nuclei classification. In [5], Malon et al have trained a deep CNN classifier to classify mitotic and non-mitotic cells. Zhang et al [8] naturally integrate deep learning and transfer learning into a single framework for classifying cervical nucleus. In [9], Xu et al have proposed a novel stacked sparse autoencoder framework consisting of two sparse autoencoder layers followed by a softmax classifier to distinguish between nuclear and non-nuclear patches. Other efforts include [10] has proposed a novel neighboring ensemble predictor coupled with CNN to more accurately predict the class label of the nucleus.