Research On Segmentation Of Breast Medical Image Based On GAN

With the rapid development of medical technology,breast cancer still has a high incidence and is the most common malignant tumor in women worldwide.According to statistics from the National Cancer Center,the incidence of breast cancer in China is between 40 and 50 per 100,000,which is three times higher than 20 years ago.Therefore,breast cancer is one of the problems that the medical profession needs to continue to overcome.In recent years,the mortality rate of breast cancer has shown a downward trend.One of the most important reasons is that the development of early screening for breast cancer has won enough time for the treatment of breast cancer,which can prolong patient’s life and even cure.Mammography is the most basic and commonly used method for detecting early breast cancer,and the diagnosis can be diagnosed by analyzing mammography images.Computer aided diagnosis promoted by artificial intelligence has also played a relatively important role in clinical medicine.Accurate breast mass segmentation can better facilitate disease identification and classification,which is a key step in computer aided diagnosis.As for the segmentation of breast mass,the methods can be divided into traditional machine learning methods and deep learning methods,both of which have their own advantages and disadvantages: traditional algorithms are simple to operate,but manual feature extraction has limitations;deep learning algorithms can make full use of images features,but complex convolution calculations will cause loss of pixel information.Through research and analysis of existing technologies,the main contents of this article are:(1)For pre-processed mammography images,this paper uses a multi-scale full convolutional network segmentation model M-FCN,and uses four full convolutional networks with different convolution kernel settings to operate in parallel.The segmentation method based on multi-scale full convolutional network is superior to the traditional process of manually extracting features.The operation of the convolution kernel can extract more important and difficult to find pixel information;it is also better than a single full convolutional network and can extract multi-scale mammography image pixel information.Aiming at the proposed M-FCN segmentation model,a further segmentation optimization model M-DFCN adopting the principle of dilated convolutions is proposed.Under the effect of dilated convolutions,reduced the loss of pixel information caused by continuous convolution pooling layers during the calculation of full convolutional networks.Can expand the receptive field while reducing the loss of pixel information.(2)Drawing on the game theory of generative adversarial network(GAN),this article further proposes an optimization method M-DFCN-A for breast mass segmentation of mammography images.Through adversarial training,the breast mass segmentation method M-DFCN is made a higher order constraint judge.The training strategy of adversarial training is explored,and the performance of the segmentation method under fast alternating,slow alternating,and dynamic alternating schemes is analyzed.Experiments show that the breast mass segmentation method M-DFCN-A-dynamic proposed in this paper is aimed at the mammography image data sets DDSM-BCRP and INBreast.Under the optimal strategy,the Dice coefficients of breast mass segmentation are 91.88% and 91.2%.Compared with the existing breast mass segmentation methods,the method in this paper has higher accuracy.