Research Of Prostate MR Image Segmentation Based On ResU-Net Deep Network

Prostate diseases(such as prostate cancer,prostate hypertrophy,prostatitis,and benign prostatic hyperplasia)are a common category of male diseases.The diagnosis,treatment and prognosis of these diseases are very dependent on the MR image analysis of the prostate.Accurately segmenting the prostate area in the MR image of the prostate is the key step in the above process.However,due to the large deformation of the prostate tissue and the blurred surrounding contours,manual image segmentation by the imaging doctor is time-consuming and laborious,and is accompanied by limited reproducibility.Therefore,there is a great clinical need for efficient and accurate automatic segmentation methods.In recent years,with the development of deep learning,deep learning has made significant progress in the field of image classification.At present,there have been examples of applying deep learning to prostate MR image segmentation,but it is limited by few training samples and categories.The problem of imbalance leads to an unstable training process,and the designed network is not optimized for prostate MR images,resulting in insufficient accuracy,and the lack of preprocessing steps for prostate MR images all bring difficulties to develop deep learning in the problem of prostate MR image segmentation.To address these issues,this paper proposes a deep learning-based prostate MR image segmentation algorithm.By decomposing the segmentation problem of prostate MR images into three stages: image preprocessing stage,network training stage,and network inference stage,low coupling and high cohesion of the overall algorithm are achieved.In the image preprocessing stage,this article will focus on the image itself by analyzing the image modalities,voxel spacing,size and other attributes,the image will be unified in format,size cropping,voxel value truncation,resampling,normalization,and the segmented network hyperparameters will be designed based on the analysis;in the network training stage,this paper uses the network Res U-Net combining U-Net and Res Net to achieve accelerated convergence of the network;at the same time,for the problem of category imbalance,an oversampling strategy is designed at the source level to balance the proportion of positive and negative samples,which alleviates the problem of category imbalance in the network input;at the loss function level,the loss function of cross entropy plus Dice coefficient is used to make the network pay more attention to the foreground area;In the network inference stage,by improving the original sliding window strategy,the original three-layer cycle is reduced to two-layer cycle,improving GPU utilization,thereby speeding up the inference speed,and designed a Gaussian template for output with the network multiplying reduces the problem of inaccurate edge segmentation due to excessive zero-padding inside the network.This paper compares with other methods in this field on multiple data sets,and makes a full ablation analysis of itself.Through experimental demonstration,the effectiveness of the proposed method is confirmed.