Cervical MR Segmentation Based On Non-rigid Registration And Convolutional Neural Network

Tumors have always been a disease that threatens the health of human life,and the incidence of tumors has continued to rise in recent years.Radiation therapy is an effective medical treatment for tumors,and precise location of lesions is the key to radiation therapy.Cervical magnetic resonance(MR)images are commonly used in clinic,which can well show the cervical tissue and its surrounding tissue structure.However,it is necessary to segment the cervical MR image to obtain the area of the lesions for radiotherapy.At present,cervical MR image segmentation mostly adopts manual mode,and the workload is large.The segmentation result is reproducible.Aiming at this problem,this paper adopts a way based on non-rigid registration and convolutional neural network model to segment the cervical MR images,and has achieved certain results.The main content of this article can be summarized as the following two aspects:1.For the traditional mutual information measurement method,the spatial characteristics between the gray information of the image are neglected.In the case where the anisotropy of the pixel gray is high,the effect is not ideal.In this paper,self-similarity weighting is used to integrate into multiple features Mutual information.Firstly,the multi-dimensional feature metrics incorporating gradient information and gray information are constructed by the K nearest neighbor calculation method.Then,the partial derivative formula of the multi-dimensional feature metric for the registration deformation parameter is deduced,and the gradient descent method is used to optimize the image.The global affine transformation combined with the local B-spline transform completes the non-rigid registration of the image and obtains the result of the coarse segmentation of the cervical MR image.2.For the complexity of medical images,the traditional segmentation algorithm is insufficient for the requirements.In this paper,the gold standard image obtained by doctors is employed by the symbol distance function to plan the target segmentation area.In the planned area,the seed points are sampled by random sampling,and the image patches are obtained centering on the seed points.By training a large number of image patches in the VGG-19 network that has been pre-trained,the convolutional neural network model required for this paper was established.The coarse segmentation result obtained by registration is treated as a seed point in the specified area after the processing of the symbol distance function.The image patch required for the test is also centered on the seed point and placed in the trained position.In the convolutional neural network,the predicted segmentation image of the network is obtained.Finally,the image segmentation results will be obtained after later image processing such as image denoising,curve fitting.The experimental results show that the proposed method has a certain improvement in accuracy compared to the registration method.