Statistical Feature Method Study Of Prostate Magnetic Resonance Medical Image Analysis

With the wide application of medical imaging technology,the study of medical image analysis has attracted more and more focus.Stable and reliable prostate tissue segmentation is of great significance in many field such as: analyzing abnormal tissue pathology structure,realizing early stage computer aid diagnosis,protocoling therapeutic schedule,surgical implementation and tracking treatment effect.As the anatomical information,which contains shape,location and size,of prostate tissue is different in various individual or various stage of same individual.The bias field which is produced by the synthetic action of imaging device and patient-specific is normal in magnetic resonance image.The low-resolution is also general in medical image.Due to the influence of bias field effect,noise,irregular tissue motion,low-resolution,the edge of different organ is usually indistinct.Consequently,accurate prostate tissue autosegmentation or semiauto-segmentation is still a challenge task.In this background,we deeply investigate the image enhancement method aiming medical application and prostate tissue segmentation method in this paper.Some efficient methods have been investigated,such as superficial neural network based medical image super-resolution reconstruction method,histogram statistical analysis based magnetic resonance image bias field correction method,hierarchical prostate magnetic resonance image segmentation via variation level set with shape prior.The main works and innovation point is summarized as follow:(1)Aiming to the normal low-resolution of medical image,this paper studies a superficial neural network based super-resolution reconstruction method.In this paper,we employ a convolutional neural networks for the reconstruction of super-resolution image basing the general image degradation model,which learns the nonlinear mapping from the low-resolution space to high-resolution space directly.The loss function which means the mean squared error of reconstructed high-resolution image and the original high-resolution image is applied to guarantee the effectiveness of training model.We use Randomized Rectified Linear Unit to solve the problem of over compression.And the Nesterov′s Accelerated Gradient method is used to accelerate the convergence of loss function and avoid the large oscillations.(2)To solve the bias field problem caused by the synthetic action of imaging device and patient-specific in magnetic resonance image,we put forward the histogram statistical analysis method for bias field harmonic approximation.By constructing an energy function basing the generally used magnetic resonance image model,the bias field correction is transferred as the minimization of energy function.According to the optimization of energy function embedded physical property of true tissue,we can obtain the correction of bias field.(3)Aiming to the variation shape,uncertain location,weak edge problem of prostate segmentation,this paper propose a variational level set model based hierarchical scheme to realize tissue segmentation.Firstly,the tissue localization is used to simplify the background.Then,we formulate the level set function combined shape prior to segment the tissue automatically.The shape prior of prostate tissue is applied as the penalized term to ensure the accuracy of segmentation.A large set of experiments demonstrates the superior advantages of our approach on both segmentation accuracy and noise sensitivity comparing to the state-of-the-art approaches.