Super-resolution Analysis Of Magnetic Resonance Imaging Based On Generative Adversarial Networks And Its Application In Molecular Pathological Information Prediction Of Breast Cancer

Breast cancer is the most common malignant tumors among women in the world with the increase of its incidence and mortality.How to use the molecular pathological information of tumor to improve the early detection and treatment of breast cancer is an urgent task.Both histological grade of breast cancer and Ki-67 provide significant guidance information for the prognosis and treatment of breast cancer.Magnetic resonance imaging(MRI)is a multi-parameter imaging in which the combination of dynamic contrast enhancement MRI(DCE-MRI)and diffusion weighted imaging(DWI)can improve the accuracy of cancer diagnosis.DCE-MRI can not only reflect the morphological characteristics of the lesion tissue,but also reflect the microscopic information.Its image resolution is high,but the contrast agent needs to be injected in advance.DWI reflects the biological characteristics of tumor such as vascular structure and water content,but its image resolution is low.For DWI,it is not easy to obtain images with ideal resolution due to technical limitations.Low-resolution(LR)images often hamper the accurate diagnosis of a professional physician.Super-resolution(SR)technology is developed to solve this problem of LR images by generating high-resolution(HR)images from LR images.The purpose of this study is to obtain the SR image of DWI through SR technology,and then predict the histological grade and Ki-67 expression of breast cancer based on the SR image characteristics of the lesion area of DWI.The research emphasis of this paper includes the following two parts:(1)Super-resolution image reconstruction: The patient information of 322 patients with invasive breast cancer was systematical y reviewed and randomly divided into training set and testing set,including 222 people in the training set and testing set of 100 people.Due to generative adversarial network(GAN)is a powerful the generative network,so this study adopted two kinds of method of super-resolution reconstruction based on GAN: SRGAN and its variants EDSR.This study used the LR and HR images of DCE-MRI in training set to train SRGAN and EDSR.The reconstruction effect of the super-resolution model was verified by using the HR,SR and bicubic interpolation(BI)images of DCE-MRI in validation set to calculate the Peak Signal Ratio(PSNR)and Structural Similarity(SSIM).The PSNR obtained by the SRGAN method is 1.23 d B higher than the PSNR obtained by the bicubic interpolation method,and the SSIM is higher than 0.008.The PSNR obtained by the EDSR method is 3.243 d B higher than the PSNR obtained by the bicubic interpolation method,and the SSIM is 0.018 higher.It can be seen that the quality of SR image obtained by the two super-resolution reconstruction methods is higher than that obtained by the bicubic interpolation method,and the quality of EDSR image is the best.ADC images were calculated according to the b value of DWI images,and the original LR images of ADC in the testing set were used to obtain the corresponding SR images by using SRGAN and EDSR.(2)Prediction study of histological grade and Ki-67: After obtaining 3D lesions of DCE-MRI images by semi-automatic segmentation,the lesion areas of ADC original LR images,bicubic interpolation(BI)images and SR images can be obtained by image registration.Then,statistical,morphological and texture features were extracted from the lesion regions of different ADC images.Univariate analysis and multivariate analysis were performed on the prediction tasks of histological grade and Ki-67 expression.Logistic regression method is used in univariate analysis;genetic algorithm is used to obtain the optimal feature subset and logistic regression method is used to predict in multivariate analysis.Then AUC,ACC,sensitivity,specificity and other indicators were calculated to evaluate the prediction model.In this study,the histological grade and Ki-67 expression of breast cancer were predicted based on the image characteristics of the lesion area of ADC original LR image,BI image and SR image.The experiment showed that compared with the original LR image or BI image of ADC,the SR feature set extracted by SRGAN or EDSR can obtain better prediction results.When multivariate analysis was carried out,the optimal AUC of histological grading prediction study was obtained from the image characteristics of SRGAN,and the optimal AUC was 0.831.The optimal AUC of Ki-67 prediction study was obtained from the image characteristics of EDSR,and the optimal AUC was 0.864.The results showed that the super-resolution reconstruction technology could improve the image quality of ADC,and the three-dimensional lesion characteristics of ADC SR image could better improve the prediction performance of prediction model for breast cancer histological grade and ki-67 expression,and provide more accurate guidance for diagnosis and treatment of breast cancer...