Research On Deep Learning Based Segmentation Of Prostate MR Image

Prostate cancer is one of the most common malignant tumors threatening the health of men.Radiotherapy is the mainstream treatment of prostate cancer.Its basis and premise operation is to accurately locate and segment the lesion and normal tissues around it.Magnetic resonance imaging(MRI)has become one of the most effective assistant tools for the diagnosis of prostate diseases in modern medicine because of its non-contact and painless advantages.The prostate is small,the contrast of the region of interest(target area)around it is low,and the boundary between organs and tissues is blurred,which makes it difficult to distinguish.In addition,there are individual differences in the structure and pathological changes of the prostate in each patient,which makes it difficult for doctors to diagnose and analyze.Manual segmentation of target area is a common method in clinic at present,but the segmentation results are greatly influenced by doctors'subjective opinion.Besides,manual segmentation is really time-comsuming and has efficiency limitations Therefore,it is of great importance for the diagnosis and clinical application of prostate cancer to develop an image processing algorithm that can accurately and automatically segment prostate in MRI.Nowadays,in the context of the big data era,deep learning has gradually evolved into an important research direction in the field of artificial intelligence,and has been widely used in medical image processing.In this thesis,convolutional neural network is applied to medical image processing by using deep learning technology.Three kinds of prostatic MRI segmentation models based on deep learning are proposed,which can greatly improve the accuracy of segmentation on the basis of traditional methods.The first prostate MRI segmentation model,PSP net,is proposed in this paper.PSP-Net uses the improved residual network to extract the global feature map.On the basis of the feature map,it applies the idea of hierarchical feature extraction in the pyramid scenario model,extracts the hierarchical features with convolution kernels of different sizes,and finally fuses these hierarchical features with global features to obtain a more robust feature map.Based on U-Net network,two prostate segmentation models are proposed based on improved U-Net network.Inspired by the superiority of residual structure,Res U-Net introduced residual structure on the basis of the original network to construct a deeper network layer segmentation network,and introduced residual structure in the direction of expansion path.While the number of network layers increased,Res U-Net still maintained a higher training efficiency,so as to improve the segmentation accuracy.Considering the 3D characteristics of MRI data,3D U-Net expands the 2D deep learning network to 3D deep learning network,so that the network can directly process 3D MRI data.The model starts from different paths,and each path has different perceptual horizons.It extracts the same feature in different degrees,and strives to obtain more detailed and global information of the feature map.The experimental results show that the proposed segmentation method can significantly improve the accuracy of prostate segmentation in MRI,and the segmentation accuracy reaches 89.98%,which better realizes the automatic segmentation of prostate MRI image.