Deep Supervision Combined With Block Learning To Detect The Nucleus Of Breast Pathological Images

Early diagnosis and treatment of breast cancer can effectively reduce the mortality of patients.Histopathological image analysis plays a key role in the diagnosis of various cancers.For the quantitative analysis of histopathological images,the features are usually quantified according to the appearance and shape of individual cells and the distribution and arrangement of cell clusters.Therefore,automatic nuclear detection for single cell recognition and localization is an important task in many pathological analysis.However,automatic kernel detection needs to identify a large number of kernel objects in a large number of digital slices,and the variability of kernel size,shape and texture and partial kernel overlap make the detection task timeconsuming and laborious,and manual processing is easy to make mistakes,some traditional target detection methods have some problems,such as relatively low test recognition accuracy and weak generalization ability,which cannot achieve satisfactory performance.In this paper,the method of deep supervision combined with image block learning is used to detect the nucleus of breast pathological image for computer-aided diagnosis and treatment.The experimental results show that the method in this paper has better performance than the previous deep network method on the data set of H & E stained histopathological images.The main research work and achievements of this paper include:(1)The model proposed in this paper uses convolutional neural network as the backbone network,and adds channel attention after down sampling in each layer of the model backbone network.Each channel represents a special detector to detect the importance of each characteristic channel,enhance or suppress different channels,and guide the computing resources to the part with the largest amount of information in the input signal.Combined with the deep supervision mechanism,using multi-scale and multi-level feature learning,each network layer is responsible for generating a certain scale prediction,providing a rich hierarchical representation for the network,and determining the candidate core and its location through the feature fusion layer.(2)This paper creatively trains a new image block sampling method to sample positive and negative samples to obtain a richer sample data set.In addition to the core category label,the position expression is also added to the image block information by retaining the offset value to participate in the position regression;By training image blocks instead of the traditional pixel by pixel sliding window strategy,the network prediction speed is greatly accelerated.(3)A thorny problem of traditional kernel detection methods is how to deal with multiple candidates of each object obtained from the model.Compared with traditional kernel detection methods,a thorny problem is how to deal with multiple candidates of each object obtained by the model.In this paper,soft non maximum suppression(soft NMS)is used to attenuate the detection score into a continuous function of IOU,rather than setting the score to zero.LPN network combined with soft non maximum suppression method can effectively reduce false negatives in target detection and significantly improve the accuracy of nuclear position prediction.The optimal detection accuracy of this network model in the dataset provided by Case Western Reserve University is 86.9%,which is improved and has great advantages compared with other models.In order to further verify the effectiveness of the network model in this paper,we compare the performance with other network models through a variety of evaluation criteria,proving the feasibility and superiority of the deep supervision combined with image patch learning method for detecting nuclei in breast pathology images.