Research On Breast Nodules Segmentation Algorithm Based On DCE-MRI Image

Magnetic resonance imaging（MRI）is an important means of breast disease examination.Different MRI modal images can reflect different physiological characteristics.Breast nodule is an important disease of many common breast diseases.Doctors can obtain an important basis for clinical diagnosis by segmenting breast nodules on MRI images.In a variety of MRI modes,dynamic enhanced magnetic resonance imaging（DCE-MRI）images can clearly show the morphology of breast nodules and the structural relationship with surrounding tissues.Doctors can manually segment DCEMRI images to establish the gold standard of nodule contour.T2-MRI is a more common MRI modality in clinic,but it is difficult for doctors to accurately outline breast nodules directly on T2-MRI images.Automatic segmentation algorithm can improve the efficiency of breast nodules segmentation in MRI images and realize breast nodules segmentation in multimodal MRI images.The first research content of thesis is to realize the breast nodules segmentation algorithm of DCE-MRI image.According to image statistical analysis,17280 DCE-MRI images were selected from 72 case data to generate a data set.Taking the breast nodules template outlined by the doctor as the gold standard,multiple deep convolution neural networks of U-Net series were used for breast nodules segmentation experiments.In order to eliminate the interference of redundant parts in DCE-MRI images,a focus segmentation idea of "locating breast tissue first and then local nodules" is proposed in thesis.Based on this idea,a simple and effective breast tissue segmentation algorithm is established to reduce the image calculation area from breast,chest,shoulder and arm to breast.In order to facilitate neural network to capture effective features,a difference image reflecting the dynamic change of DCE-MRI image is proposed in thesis.The application of breast segmentation and difference image greatly improves the segmentation accuracy.A hybrid loss function is also proposed to improve the segmentation accuracy of highly difficult DCE-MRI images.The average Dice validation score of the final model in the 5-fold cross-validation is 0.8282.The second research content of thesis is to realize the breast nodules segmentation algorithm of T2-MRI image.Since it is difficult to achieve accurate segmentation directly in T2-MRI,a more popular clinical MR image modality,thesis proposes the solution of "first modal registration and then contour mapping",that is,first realize the registration of DCE-MRI and T2-MRI through the registration algorithm,and then map the breast nodule region of DCE-MRI image to T2-MRI image.In order to obtain fine registration,affine transformation is used for global alignment,and then the image is locally cut for depth transformation estimation.In view of the large difference of pixel distribution between the two modes of MRI,the normalized cross-correlation loss is used for depth transformation estimation.Before registration,local histogram equalization,contrast enhancement and denoising are performed on T2-MRI images,which greatly optimizes the image quality and registration results.In addition,based on the above research content,thesis designs a breast nodule segmentation software of MRI image,and designs multiple functions such as region of interest scaling and feature extraction for the software,which is convenient for clinical application.