Research And Implementation Of Segmentation Algorithms For Prostate Nuclear Magnetic Resonance Images

Prostate cancer is a common malignant tumor in middle-aged and elderly men.Magnetic resonance imaging(MRI)is considered to be the best medical imaging for the diagnosis and adjuvant treatment of prostate cancer because of its high resolution of soft tissue,multi-parameter imaging and scanning of arbitrary sections.Accurate segmentation of prostate from prostatic magnetic resonance images is of great value in assisting the diagnosis of prostate cancer.Although there are many segmentation methods for prostatic magnetic resonance images in recent years,these methods still can not adapt to different image data sets.To a certain extent,the success of medical image algorithm depends on high quality input features.However,feature engineering requires a lot of time and effort,and the ability of artificial feature selection is poor in expression and generalization,which can not meet the current needs of image segmentation.The multi-layer structure of deep neural network can effectively express complex functions,so that it can learn image features with strong representation ability and improve the accuracy of image recognition.In this paper,three full-convolution neural networks U-Net,V-Net and Dense V-Net are used as basic contrast schemes for prostate magnetic resonance image segmentation.Then,based on Dense V-Net and V-Net,a dense-connected full-convolution neural network FC Dense V-Net is proposed.The connection between convolution layers in FC Dense V-Net uses dense connection in Dense VNet,which can achieve a greater degree of feature reuse and effectively solve the disadvantage of existing models that can not obtain different levels of features at the same time.At the same time,symmetric deconvolution and convolution operations in V-Net are used in FC Dense V-Net for upsampling,so that deconvolution operations in FC Dense V-Net will have corresponding results.It is very important for medical image,which has a wide range of gray values,to extract deeper feature information by convolution operation,so that the network model can extract more feature information and realize the refinement of segmentation boundary.To a certain extent,it alleviates the over-fitting problem of network model caused by the small medical image data set and improves the classification performance of the model.The experimental results on the real data obtained from the clinical center of cooperative hospital show that the performance of FC Dense V-Net network model is better than Dense V-Net and V-Net.This algorithm evaluates the performance of network segmentation by four evaluation indexes: the Dice coefficient(DSC),absolute relative volume difference(aRVD),average boundary distance(ABD)and Hausdorff distance(HD).The similarity coefficient of FC Dense VNet is 89.1%,which is higher than 87.8% of Dense V-Net and 87.6% of V-Net.At the same time,FC Dense V-Net performs better than Dense V-Net and V-Net in other three indicators.