Research On Semantic Segmentation Method Of Prostate MR Image Based On Convolutional Neural Network

Prostate cancer is one of the most common malignant tumors among men in Europe and America,and its incidence is the highest among all male cancers.In recent years,the incidence of prostate cancer in China has also shown a significant upward trend.MR imaging has become one of the most effective assistant tools for cancer diagnosis in modern medicine due to its advantages of non-contact and painless.For prostate,because of its small size and concealed location,it is difficult to observe in MR images,which makes it difficult for doctors to accurately diagnose.Therefore,it is of great significance for the diagnosis and clinical application of prostate cancer to study an effective prostate MR image processing algorithm to separate the prostate from the complex background.In order to realize automatic segmentation of prostate MR images,the PSSNet(Prostate Semantic Segmentation Net)based on convolutional neural network is proposed,which consists of encoder and decoder.In the image coding stage,firstly,ResNet-101 is used to extract the features of the image,and then DDSPP(Dense Dilated Spatial Pyramid Pooling)module is designed to extract the multi-scale semantic information of the image from the features.In the image decoding stage,in order to increase the feature information,the DP(Decoder Processing)module is designed to extract the features of different layers directly from the encoder,and then the extracted features are fused with the output of the corresponding decoder.Finally,the complete information of the prostate was gradually recovered by up-sampling.In the training stage,in order to improve the accuracy and generalization ability of the model and prevent over-fitting,it is necessary to pre-process the original data sets.Firstly,in order to reduce the impact of invalid samples on training results,the existing MR image data sets of prostates are optimized by deleting the images that do not contain prostate structure.Then,the optimized prostate MR image data sets are expanded by adding Gauss noise and enhancing image contrast.Finally,PSSNet is trained by using the expanded data sets.PSSNet was tested on the optimized and expanded MR image datasets of prostate.The segmentation accuracy of PSSNet was over 99%,the DSC(Dice Similarity Coefficient)was 0.954 and the HD(Hausdorff Distance)was 0.876mm.The experimental results show that the proposed method has high segmentation accuracy and robustness for prostate MR image segmentation.