The Classification And Detection System Of Cervical Cells Based On Improved SSD Network

The manual screening methods for cervical cancer have some problems,such as high cost,low efficiency.And its reliability and accuracy are greatly affected by doctors’subjective interpretation.With the development of computer science,image processing and deep learning are gradually applied in cells’ recognition and detection.However,the current researches on the classification of cervical squamous epithelium are insufficient,and the classification results are simplified.At the same time,there are also many problems,such as low accuracy and weak generalization.Especially,the recognition accuracy of highly pathological cells needs to be improved.To solve these problems,an automatic screening system for cervical cancer cells based on improved SSD(Single Shot MultiBox Detector)network is proposed in this paper.To obtain images,the system can automatically focus and scan samples with the microscope,then classify and detect the cells.The cervical squamous cells can be divided into normal cells,LSIL(Low-grade Squamous Intraepithelial Lesion),HSIL(High-grade Squamous Intraepithelial Lesion)and squamous cell carcinomas.The main works of this paper are as follows:(1)Build a microscope imaging platform.We set up a row-by-row scanning on X axis and Y axis,and design automatic focusing algorithm on Z axis.For the function of image’s sharpness evaluation,we fuse the traditional Laplacian operator and local variance information,and the searching algorithm of focus plane is designed by combining the hill-climbing search and the function approximation algorithm.(2)The cervical squamous cells were classified and detected by deep learning technology.On the basis of the original SSD network model,the feature map is fused by forward and backward feature,to improve the disadvantage of the original network which is not sensitive to small object detection.We use the idea of bilinear convergence feature analysis to adjust the network structure.For the calculation of the loss function,the central loss is added in this paper,which is more suitable for the phenomenon that intra-class difference is greater than inter-class difference in cells’ classification.(3)To verify the automatic scanning platform and the cells’ classification and detection algorithm,we design the experiment.The result shows that the improved focusing algorithm can obtain high-quality images more quickly,and the whole platform can automatically complete the scanning of the samples.The improved network model’s mAP(mean Average Precision)is 81.53%,which is 7.54%higher than YOLO(You Only Look Once)algorithm and 4.92%higher than SSD algorithm.Finally,we summarize the whole paper and point out the future research directions.There are 44 figures,10 tables and 52 references in this paper.