Research On Automatic Segmentation Of Prostate MRI Images Based On Deep Learning

The prostate gland is the accessory gland of the male reproductive system.Benign prostatic hyperplasia,prostatitis and prostate cancer are the most common prostate diseases.The clinical manifestations,physical signs and MRI results of these three diseases are relatively similar,and it is not easy to distinguish.According to a report released by the International Agency for Research on Cancer under the World Health Organization(WHO),the number of prostate cancer patients in the world reached 1.414 million in 2020,accounting for 14.1 percent of the total,making it the second most common male cancer after lung cancer.The inertia of prostate cancer is high.If it can be found and treated as early as possible,the treatment difficulty and fatality rate of prostate cancer will be reduced.Accurate segmentation of the edge of the prostate region in pelvic magnetic resonance imaging(MRI)images is very important for the diagnosis of prostate diseases and plays an auxiliary role in the screening of prostate cancer.Due to the relatively small proportion of the prostate region in pelvic MRI images,the blurred boundary between the prostate region and other areas,and the great differences in the size,shape and position of the prostate region in different patients,it is very difficult to accurately extract the boundary information of the prostate.Currently,the segmentation of the prostate region depends on the radiologist manually marking the prostate boundary in the pelvic MRI image,which is a very time-consuming and labor-intensive process.Therefore,the automatic segmentation method of prostate MRI image has clinical significance and research value.This topic will be based on deep learning technology to carry out research on automatic segmentation of prostate MRI images.Firstly,a deep learning framework(LDU-Net)based on improved U-shaped convolutional neural network(U-Net)is proposed.This novel prostatic automatic segmentation network uses U-Net as the main framework of the segmentation task,and introduces the parallel dilated convolution attention mechanism(LD module)into the coding and decoder of the framework,so as to enhance the attention of LDU-Net on quantified information and strengthen the segmentation ability of the segmentation network at the edge of the prostatic region.Secondly,in view of the differences in the size,shape and position of the prostate region of different patients,deformable convolution(DCN)was introduced on the basis of LDUNet to obtain a new segmentation network,LDDU-Net,which avoided the overfitting problem in the process of model training and improved the perception ability of segmentation network.Thus,the generalization ability of the network is enhanced.The results showed that LDDU-Net had excellent performance in segmental prostatic region,and the average Dice similarity coefficient was 90.93%.Therefore,LDDU-Net has high segmentation accuracy and can complete the task of automatic segmentation of prostate MRI images.