Research On Circulating Tumor Cell Detection Algorithm Based On Convolutional Neural Network

Worldwide,the incidence of malignant tumors has increased year by year,seriously jeopardizing the physical and mental health and life quality of patients.The detection of circulating tumor cells in peripheral blood is an effective early diagnosis method for tumors,and is also of great significance for the late treatment of malignant tumors.Microscopic image analysis is an important branch of medical image processing.The detection and quantitative analysis of cells based on high-throughput microscopic images using image processing and machine learning method has long been a hot topic.Based on characteristics of circulating tumor cell images and the current fast developing convolutional neural network method,main contributions of this thesis include the following aspects:1.A system for locating suspicious circulating tumor cell regions based on morphological method is proposed.In order to make full use of the optical information of blood,the microscopic images are scanned at ten different focal lengths to obtain ten images.First,the original image is transformed into gray images and binarization is performed.Next,morphological method is used to select suspicious circulating tumor cells regions.Finally,detected overlapping suspicious cells regions are combined to give the final suspicious region.We also evaluate the image quality of all ten layers of detected area to select the best layer to store in the database.Related information such as the location of the suspected cells and the size of the suspected cell area will be stored in the system for better clinical diagnosis in the next step.2.A circulating tumor cell detection algorithm based on CNN network structure is proposed in this thesis.The proposed CNN structure is based on the improved ZF network model to achieve the classification and detection functionality of circulating tumor cells.First,we collect and organize the CTCs sample image database,and then we segment the whole slide images into image size of 2048? 2048.They are further divided into 82? 82 image patches to go through the CNN network to be classified into positive and negative samples.Finally,Under the guidance of experienced doctors,data labeling and CTCs sample database were re-established.3.A circulating tumor cell detection algorithm based on Faster R-CNN network structure is proposed.The RPN network is added after the basic convolutional neural network,and the main function is to extract the preferred area.Then,the ROI pooling layer integrates the features of the corresponding pre-selected areas into fixed-length feature vectors,and then feed forward to the softmax classification layer and also the bounding box regression layer to obtain the final target class and precise target bounding box position for accurate identification and detection of circulating tumor cells.According to the best of our knowledge,this has been the first time that the Faster R-CNN network structure is used for circulating tumor cells detection.High precision is achieved in circulating tumor cells detection and identification.