Research On Cervical Cancer Cell Detection Technology Based On Deep Learning

Cervical cancer is a gynecological disease with a high fatality rate,which seriously endangers women’s health.At present,the core technology of the processing method for cervical cells mainly includes three major steps,namely cell segmentation,feature extraction and feature fusion,and cell classification.The performance of traditional cervical cell processing methods largely depends on the accuracy of segmentation and the effectiveness of hand-made features,and the accuracy of any step will affect the final detection effect.The object detection algorithm combines object segmentation and recognition into one,and does not require any pre-segmentation steps.The final detection effect will be good as long as the detection algorithm is effective.In addition,deep learning has very good application prospects.The research of cervical cancer cell detection in this article is based on the object detection algorithm of deep learning.In view of the superiority of Retina Net object detection algorithm in the field of object detection,this paper take it as a basic algorithm for research.Firstly,add a multi-scale convolution structure to the basis of the Retina Net algorithm to extract more size object features,avoid incomplete object feature extraction,and optimize the detection efficiency of the algorithm.Secondly,data augmentation technology is used to increase the diversity of experimental data and to make the model more applicable in more scenarios.Finally,the transfer learning method is used during experimental training to improve the training speed of the model.The experimental results show that the detection accuracy of the cervical cell image data set by the Retina Net algorithm with a multi-scale convolution structure has achieved a 61.58% m AP,which is 3.21% higher than the original Retina Net algorithm(m AP=58.37%).On the basis of adding a multi-scale convolution structure,this article introduces the SE module to make the model pay more attention to channel features with large amount of information by assigning weights to feature channels,while suppressing those unimportant channel features,The detection accuracy was further improved,and63.40% m AP was achieved.In order to verify the superiority of the improved algorithm,this paper finally conducted experiments on YOLOv4 and Faster R-CNN.The m AP values of the two are 58.40% and 55.28%,respectively.It can be seen that the algorithm proposed in this paper has more advantages in detection accuracy.