Prostate MR Images Segmentation Based On Deep Learning

Prostate cancer is the most common cancer in elderly men.Prostate images obtained by magnetic resonance imaging have been widely used in the diagnosis and treatment of prostate cancer.The doctor needs to accurately segment the area of the prostate tissue in the MR image before diagnosis and treatment.However,manual segmentation is time-consuming and laborious,and due to the large difference in prostate morphology and low contrast with adjacent structures,the traditional segmentation methods still have the disadvantages of low accuracy and slow speed.With the prominent advantages of deep learning technology,the methods based on deep learning have shown great potential in image processing.Compared with traditional methods,deep learning networks can extract the deep abstract features of images and realize end-to-end processing.Therefore,it is of great significance to study and design deep learning networks for prostate MR images.The main contributions of this thesis are as follows:1.This thesis briefly summarizes the research background,significance and current research status of prostate image segmentation,and makes a brief classification and summary of the existing methods of prostate image segmentation.After that,this thesis elaborates on the theoretical knowledge of image segmentation and deep learning in detail,and the classic image segmentation algorithm based on deep learning.2.In order to solve the problem of low segmentation accuracy caused by insufficient recognition of small target area by fully convolution networks,this thesis proposes segmentation network based on encoding-decoding structure with residual learning and dense upsampling convolution to perform prostate MR images segmentation.The encoding-decoding structure can recover the lost spatial information of the fully convolutional networks to a certain extent.Based on the encoding-decoding structure,residual learning is used to optimize the training of deep network and avoid the degradation of the network.In addition,in order to further recover the spatial detail information lost in the encoding process,the dense upsampling convolution is used to capture and recover the detail information,so as to improve the network's ability to recognize and recover the details of the image.The experimental results show that the encoding-decoding segmentation network based on residual learning and dense upsampling convolution can effectively improve the segmentation accuracy of prostate MR images.3.In view of the excellent performance of the generative adversarial networks in a variety of computer vision tasks and it has achieved good segmentation effects in the segmentation of natural images.In this thesis,adversarial learning is introduced to train the segmentation network.Through the mutual adversarial training between the segmentation network and the discrimination network,the output segmentation results by the segmentation network are closer to the results of the ground truth,and the segmentation performance of the network on prostate MR images is improved.In addition,the encoding-decoding structure is not enough to extract the multi-scale feature,the receptive field block is used to obtain and fuse the multi-scale information of deep features in the segmentation network,so as to improve the discriminability and robustness of features and further enhance the segmentation accuracy of the network.Therefore,this thesis proposes a segmentation network based on adversarial learning and multi-scale feature fusion.On the dataset of prostate MR images,DSC and HD achieve 89.56% and 7.65 mm respectively,which verifiey the effectiveness and real-time performance of the proposed algorithm model.