Research On Prostate MRI Segmentation Algorithm Based On Deep Convolution Network

According to the 2012 statistics of the American Cancer Society(ACS),the incidence of prostate cancer in Europe and the United States ranks first among all male malignancies,and the mortality rate caused by prostate cancer is second only to lung cancer and second only to lung cancer.Accurately segmenting prostate magnetic resonance(MR)images plays an important role in the diagnosis and treatment of prostate diseases.The lack of clear edges between the prostate and other anatomical structures makes it difficult to accurately extract boundaries.The complicated background texture and the huge changes in the size,shape and intensity distribution of the prostate itself make the segmentation more complicated.In clinical practice,medical images can usually be manually segmented by radiologists,which is an expensive and time-consuming process.In clinical practice,automatic segmentation of prostate MR images is very much needed.AI technology has been widely used in the recognition of natural scenes.In terms of medical imaging,AI technology has a wider landing scene.Through AI technology,it can effectively improve the doctor’s diagnosis and treatment rate and diagnostic accuracy,and make the analysis technology of medical images sink and shorten The waiting time for patients to see a doctor reduces the cost of seeing a doctor.The latest advances in machine learning,especially in deep learning,are helping to identify,classify,and quantify existing medical images.Deep learning can automatically summarize hierarchical features from the data,instead of manually discovering and designing features based on domain-specific knowledge as in the past.However,the characteristics of blurred borders and large changes in shape and intensity of prostate MR images are also a challenge for the design of segmentation networks.In this paper,based on the characteristics of the prostate MR image and the shortcomings of the existing segmentation methods,we propose an improved efficient and accurate prostate MR image segmentation method based on iterative positioning on the segmentation frame,network structure and loss function.This research mainly has the following three contributions:1)A 3D U-shaped network with resolution perception is proposed to balance the difference in spatial resolution between in-plane and out-of-plane.In addition,a residual structure with an instance-batch normalization(IBN)is used to improve the learning and generalization capabilities of the network.2)The Case-wise loss function is introduced to alleviate the data imbalance problem caused by individual differences in prostate MR images.This allows individuals with smaller and more difficult-to-divide areas to receive better training.3)In the inference stage,a coarse to fine prostate segmentation method with iterative localizing refinement is used.The coarse to fine segmentation method can reduce the input area to reduce the influence of complex background.The iterative positioning method we proposed can accurately locate the ROI region without the training step of the coarse segmentation model.The results verified on the MICCAI 2012 prostate segmentation challenge dataset(Promise 12)and NCI-ISBI prostate segmentation challenge dataset prove the effectiveness of our method.