Research On Diagnosis Method Of Prostate Cancer Based On MR Images

Prostate cancer is one of the main diseases of middle-aged and elderly men.Early detection and treatment of prostate cancer is the key to reducing the mortality of prostate cancer patients.Magnetic resonance imaging is considered to be the most effective method for prostate cancer imaging by virtue of its advantages of multi-directional and multi-parameter imaging.O n the basis of this imaging,research on the classification method of prostate cancer lesions based on convolutional neural networks and the development of corresponding systems can provide theoretical basis and means for the early diagnosis of prostate cancer.In view of the multi-modal characteristics of prostate MR images,this paper proposes a SERes Net18 network.The input of the model is a multi-channel ROI formed by stack ing multi-modal images.Using the Squeeze-and-Excitation(SE)module can achieve the purpose of assigning weights to the channels,thereby distinguishing the importance between modalities.AUC = 0.84;for the low resolution of the prostate MR image itself,and the resolution is exponentially lower during the continuous convolution process,a Res Net model based on FPN is proposed,which combines the characteristics of high resolution and high detail information with low The fusion of features with high resolution and high semantic information helps to improve the classification effect of small lesions.The network performance is AUC = 0.82.In order to extract the spatial features of prostate lesions,a Res Net model based on 3D multi-input is proposed.Each input corresponds to three convolution branches.After the branch features are fused,the resolution of the feature map can be maintained by the stacked hole convolution.The network performance is AUC = 0.84.Finally,Bagging’s sampling method is used to reduce the impact of data on the network,and the regression equation is assumed,and the weight of each network is obtained using the least square method.The performance of the integrated model is AUC = 0.87,which is better than the performance of any model in the PROSTATEx competition in previous years.Finally,in order to make the integrated learning model easy to use in hospital scenarios,this paper designs a cancer diagnosis system based on prostate MR images.The system can realize the functions of loading preview of prostate MR images,patient information management,image preprocessing,lesion classification and model training.