Research On Deep Learning Based Circulating Tumor Cells Detection

With the development of living condition,people pay more and more attention to their health needs.Cancer is a disease that affects human health.If it can be detected and treated in the early stages of the disease,it will greatly save patients.How to detect early and accurately in the analysis of the tumor situation of patients has become a current research hotspot.Circulating tumor cells are free in the peripheral blood,which shed from the primary tumor and enter the peripheral blood,carrying tumor related information.By capturing those cells for quantitative and material analysis,it is possible to assess the statue of patient.Due to the rarity of circulating tumor cells,in vitro detection methods,it is generally necessary to undergo enrichment and identification processes to improve detection efficiency.For the detection of circulating tumor cell images,early manual detection methods and traditional image processing algorithms are inefficient.As a current mainstream target detection algorithm,deep learning has significant advantages over traditional methods in detection speed and accuracy,so it is widely used in the field of medical image processing.Based on the theory of circulating tumor cell detection and target detection,this paper proposes the following two detection methods to realize automatic detection of circulating tumor cells,and uses real clinical circulating tumor cell images to evaluate the proposed methods.(1)RS-SVGG,a two-stage model for circulating tumor cell detection,is proposed.RSSVGG uses a multi-scale region selective search algorithm to realize regional localization of circulating tumor cells.By scaling the original image of clinical circulating tumor cells,the cells can adapt to the preset prediction frame size,solving the problem of locating large cells and cell clusters.SVGG is a classification network,based on VGG-16 network,is used to detect the final circulating tumor cells from the suspected circulating tumor cells screened in the previous stage.Experimental results show that RS-SVGG is superior to traditional image processing methods,with an accuracy of 97.56%,which is also superior to human detection.(2)RFB-YOLOv5,a single-stage model for circulating tumor cell detection,is proposed.Based on the YOLOv5 model,tumor cells can be located and recognized directly on the original image of circulating tumor cells.Because the existing clinical samples of circulating tumor cell data are insufficient to train the model completely,this paper proposes a method to expand the circulating tumor dataset by copying and pasting.Aiming at the problems that the sizes of circulating tumor cells are small and target information is easy to lose,a receptive field module is introduced to increase the size of the model receptive field.Hence,the small target feature extraction ability of YOLOv5 can be improved,as well as the accuracy of circulating tumor cells detection.The experimental results show that the RFBYOLOv5 model can achieve a recall rate of 99%and a rate of 83%mAP@0.5.Compared to the original YOLOv5 model,the algorithm proposed in this article significantly reduces the missed detection of circulating tumor cells and improves the prediction confidence of circulating tumor cells.