Mitosis Detection In Breast Cancer Based On Convolutional Neural Network

Accurate assessment of the degree of deterioration of breast cancer plays a crucial role in making medical plan,and the important basis for disease assessment is the number of mitoses in a given area of the pathological image.However,due to the size of the pathological image,multiple cellular and mitotic morphology,artificial mitosis detection is cumbersome and time-consuming,and the computer-aided system can reduce the subjectivity of detection and reduce the workload of medical experts.Most of the current mitotic detection models are based on the manually feature extraction method,but it is difficult to extract effective features due to the indistinctive mitotic characteristics.In order to solve this problem,based on deep convolutional neural network(CNN),this thesis designs a mitotic detection model for automatically feature extraction and conducts simulation experiments.In view of the characteristics of deep convolutional neural network and mitotic detection tasks,a mitotic detection model with rotation invariance,multi-scale and multi-level fusion features and context information was designed.Analyze the working mechanism and structural features of convolution neural network;Analyze the characteristics of mitotic detection——Rotation mirror invariance,scale invariance and sparsity,as well as the difficulties——small amount of data,inconsistent image staining,sample type unbalanced.Based on this,mitotic detection model is designed.The whole model includes image pre-processing module,training sample generation module,model component design module,model structure design module,model training and optimization module.In this paper,a tissue-based staining equalization method is used and selecting samples by using CNN to establish an effective training sample set.A cyclically symmetric convolution layer is designed.A multi-level feature hybrid full convolution neural network connecting a conditional random field mitosis detection model are proposed,and corresponding training strategies are formulated.The effectiveness of the model was verified on the open source ICPR MITOSIS 2014 dataset through the CAFFE Deep Learning Platform.