Object Detection And Segmentation Of Cervical TCT Images Based On Convolutional Neural Network

People began to pay attention to health with the development of society and economy.Cervical cancer is one of the most serious diseases threatening women's health.At present,medical image analysis is the necessary to make a diagnosis.But the result depend on the doctor's experience.The original image is complex,it will take doctors a long time and energy.For example,it takes at least five minutes for analyzing a cervical TCT image.A doctor has many patients every day,it may make incorrect diagnosis.It has great significance that research and analysis of cervical TCT images for the diagnosis of cervical cancer.The accuracy of object detection and segmentation in cervical cell image directly affect the diagnosis results.This paper has done the following work :(1)To produce my own COCO data set named TCTCOCO with the guidance of doctors and inspectors.It is based on the image of cervical TCT cells produced by Thinprep cytologic test of Changsha Second People's Hospital.In order to avoid over-fitting,I expand the data set.(2)To research and experiment on Faster R-CNN and Mask R-CNN.Faster R-CNN can only perform object detection,and Mask R-CNN can not only achieve object detection but also segment the object in the object frame area to generate a mask(complete the instance segmentation).(3)To propose the DFPN which based on the feature pyramid network(FPN)in Mask R-CNN.It adds dilated convolution to reduce the loss of image information during feature extraction for improving the accuracy of segmentation.(4)There are overlapping in the cervical TCT image,Mask R-CNN use non-maximum suppression(NMS)to filter the bounding box,NMS will cause the bounding box to be deleted by mistake.I learned that Soft NMS can solve this problem.Therefore,Soft NMS is used instead of NMS to reduce the mistaken deletion of the bounding box.Through simulation,segmentation accuracy is improved by using DFPN instead of FPN and classification accuracy is improved by using Soft NMS instead of NMS.It proved the feasibility of the algorithm in this paper.