Research On Medical Cell Detection Based On Deep Learning

Cell testtechnology has become a non-negligible part in China's medical circle as China's medical system is constantly improving in current stage.Conventional image processing technology is hard to meet the requirements of experiment due to the complex feature extraction process and high complexity of traditional target testmethod,such as HOG,LBP and RCD.In recent year,deep learning has become a hot topic in the academic circle and a large number of researchers are applying deep learning technology in the field of computer auxiliary image.The existing target testmethods based on deep learning,such as Faster-RCNN,SSD,FCRN,etc have overlapping samples,poor testaccuracy,insufficient training samples,etc.This paper further studies the cell testbased on the generative adversarial network(GAN)for the above problems.The specific research contents are as follows.(1)Creation of cell database.This paper proposes two surface characteristic cell tagging methods,surface characteristic single cell nucleus tagging based on Gaussian Kernel function and single cell tagging based on the red masking.(2)Sample amplification.This paper proposes the cell amplification model based on generative adversarial network(GAN)and improves the GAN loss function for the problem of insufficient samples in the valuable large scale medical image database.(3)Cell detection.It proposes the GAN detection algorithm in order to achieve high precision cell detection for the insufficient detection precision of the complex overlapping samples.(4)Post-treatment of test image.This paper proposes a kind of non linear diffusion image hybrid filter denoising method for the noise problem in the test result.It sets up the image denoising model based on adaptive probability,judges the noise property in the filter window through the gradient and adjacent pixel information,conducts the stage filter treatment of the pixels according to the property analysis result and outputs the denoising image.(5)Constitution of cell generative detection model.In regard of the complex sample amplification,this paper proposes research work of the GAN and test combination based on generative model,so as to combine the image generation and end-to-end cell detection and save an independent detection network.Based on the existing cell database,this paper enriches the cell database and solves the difficult problem of scarce database through the sample amplification experiment.This paper proposes the GAN detection algorithm,improves the complex and overlapping sample detection precision through amplification database training verification.The experiment is superior to HOG,SSD,FCRN,etc.The detection result further optimizes the cell detection output result through the non linear diffusion image hybrid filter denoising method.It saves the detection network structure through the research work in combination with the generative detection.