Color Style Normalization Of Cervical Cell Images Based On Generative Adversarial Networks

Cervical cancer,which originates in the cervix,is the most prevalent malignancy in the female reproductive tract.Cervical cancer cytopathology assisted diagnosis system can achieve early diagnosis and treatment of cervical cancer,thus controlling the progress of the disease and improving the prognosis.Although the system is clinically effective,it still has many drawbacks.In particular,there are some problems with cervical cytology images,which affect the diagnostic accuracy of the system.First of all,the use of hardware filters in the system will cause the cervical cell images to lose some of their color,which can change the original color of the cells in the images and reduce the contrast of each part of the images,making it more difficult for manual reading and thus increasing the possibility of missing and misdiagnosis.In addi tion,the appearance of cervical cell images,such as image color,can vary during the formation process depending on the specimen preparation method,staining protocol,scanner specifications,and other factors.Therefore,the color of different cervical cell images can vary significantly.These differences not only lead to inconsistent diagnoses within pathologists,but also weaken the performance of companion diagnostic systems and hinder their use in pathology.To solve the above two problems of cervical cytology images,this paper specifies the corresponding target images and normalizes the color style of cervical cell images to the color style of the corresponding target images.The work in this paper includes the following aspects:1.A digital filter based on conditional generation adversarial network is proposed to replace the hardware filter and solve the problem that the hardware filter changes the color of cervical cell images in the system.The implementation of digital filters is based on the pix2 pix framework,which can complete the conversion of unfiltered images to filtered images.In order to obtain more image information and calculate the diagnostic index accurately,a full-scale jump connection is added to the generator of pix2 pix,and a constraint term of integral optical density is added to the loss function.The experimental results show that the digital filtering method implemented in this chapter generates images that are approximates the real filtered images,effectively realizing the digital filtering of cervical cell images,and can replace the hardware filters.2.A color style normalization network called CSNGAN is proposed to standardize the color of cervical cell images,which solved the problem that color difference between images could affect the diagnostic accuracy.Based on Cycle GAN model,CSNGAN uses a simple convolution structure to replace its original generator network,which reduces the sensitivity of the model to useless texture information.At the same time,it increases the perceptual loss and enhances the contextual information.The experimental results show that CSNGAN achieves outstanding color style normalization effect and ensures that the original structural information of cervical cell images is not affected.In the practical application of cervical cancer pathology assisted diagnosis system,CSNGAN-based color style normalization can be used as a pre-processing step to effectively improve the classification accuracy of cervical cells.3.A color style normalization network called Multi-CSNGAN is proposed to standardize the color of cervical cell images,which makes up for the shortage of the CSNGAN model that only supports a single color style normalization.The Multi-CSNGAN model combines Res Net and multiple CSNGANs,and the output of multiple CSNGANs is linearly combined with the output of Res Net as the weight to obtain the final result of color style normalization.The experimental results show that the CSNGAN model achieves significant color style normalization for each image when it is fed into multiple cervical cell images.In addition,experiments have shown that the CSNGAN model can be used as a pre-processing step to improve the accuracy of cervical cell classification in a cervical cancer pathology-assisted diagnosis system.In summary,for the different problems of images in the cervical cancer cytopathology assisted diagnosis system,this paper specifies the target images separately and proposes corresponding methods to normalize the color style of the cervical cell images to the color style of the corresponding ta rget images.The experiments show that this paper is effective in normalizing the color style of cervical cell images,which can be used as a pre-processing step in the cervical cancer cytopathology assisted diagnosis system and effectively improve the diagnostic accuracy of the system.