Research On The Segmentation Method Of Prostate TRUS Image Based On Convolutional Neural Network

Prostate diseases are common in male reproductive system.With the continuous improvement of living standards of Chinese residents,the incidence of prostate cancer is also increasing.Transrectal ultrasound(TRUS)images of the prostate are widely used in the prevention,diagnosis and treatment of prostate diseases due to their low cost,real-time,non-destructive and non-radiation advantages.Especially in the diagnosis and treatment of prostate cancer,with the increase of modern clinical applications such as brachytherapy,cancer location and biopsy needle placement,there is an increasing demand for accurate automatic prostate segmentation technology in TRUS images.However,manual segmentation is mainly used in clinical practice,which is not only time-consuming and laborious,but also depends heavily on the experience and ability of doctors.Rapid,accurate and repeatable detection of the prostate region remains a challenging issue.This paper presents a prostate TRUS image segmentation network ProNet based on convolution neural network(CNN).Using the structure of encoder-decoder,the spatial dimension of input data is gradually reduced by using the pooling layer of the encoder,the details of the target and the corresponding spatial dimension are gradually restored by using the deconvolution layer and other network layers of the decoder.In the encoder stage,the Fusion Atrous Spatial Pyramid Pooling(FASPP)is proposed.The FASPP not only enlarges the receptive field,increases the pixel density,but also does not increase the computational complexity when extracting image features using more abundant spatial relations.In the decoder stage,Hierarchical Feature Fusion(HFF)and Generate Semantic Features(GSM)are proposed.HFF can fuse multi-scale feature images and make use of the complementarity of high-level and low-level features,so that each module of decoder can use the semantic feature mapping of the previous module to correct potential error features.GSM structure can generate different scales of semantic features.Combining global image features with local image features,multi-scale semantic information is generated by making full use of image context information.Finally,through dataset augmentation and AtoS network optimization algorithm,the speed of model initialization is accelerated and the generalization ability of network model is improved.The experimental results of different methods show that the ProNet network used in this paper can effectively overcome the problems of artifacts and boundary missing in TRUS image of prostate,and accurately segment the prostate region.The average absolute distance between ProNet’s segmentation result and doctor’s gold standard is 0.360mm,hausdorff distance is 1.228mm,Dice’s similarity coefficient is 0.9608,specificity is 0.9980,sensitivity is 0.9560.Compared with the traditional method and the deep learning methods,the performance of ProNet network is superior to the comparison algorithm in terms of both the subj ective and qualitative comparison results of segmentation contours and the objective quantitative evaluation results based on contour distance and contour area,which prove the effectiveness of the proposed method.