Prostate MRI Segmentation Using Deep Convolutional Neural Networks

Wise Information Technology of 120(WIT120)aims at assisting physicians in reading and pathological analysis by means of advanced artificial intelligence technology,so as to further improve the ability of disease diagnosis and reduce the pressure of physicians' work.With the continuous development of computer vision and deep learning algorithms,intelligent medical engineering is in full swing.Prostate cancer is currently one of the major diseases that seriously endanger the health of elderly men,as well as,prostate imaging analysis has been a hot spot in medical image processing.Before the diagnosis of prostate cancer,it is necessary to segment the prostate tissue accurately,and further analyze the lesions and medical indicators to confirm the degree of the disease.Therefore,it is important to complete accurately segmentation of prostatic tissue before the diagnosis of prostatic diseases,and it is an extremely necessary step.At present,artificial segmentation of prostatic tissue requires too much time and effort,and some traditional methods,such as the method based on atlas and deformation model,have low accuracy in segmentation of prostatic tissue.They can't accurately draw a clear outline of prostatic tissue,and then can't distinguish the prostatic tissue area from the background area.It is very disadvantageous for doctors to diagnose the prostate illness and formulate the follow-up medical treatment plan.In view of the rapid development of depth learning in the field of image segmentation,we propose a prostate MRI tissue segmentation method based on deep convolutional neural networks.The main works of this paper are summarized as follows:1.Considering that it is very difficult to make progress of the traditional fully convolution and encoding-decoding network on segmentation task,this paper considers the introduction of residual basic unit,and constructs a deep residual encoding-decoding network model.The residual coding-decoding network is still a mirror structure,and the convolutional modules are all built based on the residual network.This network model not only has deeper depth than the traditional encoding-decoding network,but also introduces skip connections with different lengths,which enhances the reuse of hidden layer features.At the same time,in order to overcome the problem of lack of data and low contrast of prostatic MRI images,a series of morphological transformations are adopted.At the same time,for the problem of uneven distribution of prostate slices,this paper uses weighted cross-entropy loss function to optimize the model training.The experimental results show that the prostate segmentation of proposed algorithm is significantly better than that of the fully convolutional neural networks and the traditional encoding-decoding networks models.2.With the number of layers of convolutional neural networks increase,more and more parameters need to be learned,and the training time also increases.This paper presents a new convolutional module which combines separable convolution and dense connection.The new convolutional module greatly reduces the parameters needed to be learn in traditional convolution methods.Based on this new module,this paper constructs a neural network structure that integrates multi-scale feature information,which can effectively alleviate the phenomenon of parameter redundancy and enhance the utilization of multiple features.In addition,in order to introduce the boundary information of prostatic image,a dual CNN segmentation model is built based on multi-scale feature fusion network.At the same time,the dual CNN uses the internal and external boundary information of the prostate for model training,and at the end of the network,two-channel feature information fusion is carried out to obtain the segmentation results of the prostate.The algorithm has achieved good results in the experiment of prostate MRI data.