Neural Network For Diagnosis Of Prostate Cancer From MRI Images

Prostate cancer is devastating malignant tumor that is hard to recognize on early stages.The “gold standard” of diagnostic is MRI and further investigating of MRI scans.Artificial Neural Networks have great potential of image recognition tasks and can be used for automated diagnostic systems that can help a lot to medical personnel.In this work a Convolutional Neural Network models were developed for prostate cancer recognition based on MRI images of healthy and non-healthy prostates.The proposed structure of a convolutional multilayer unidirectional neural network consists of an alternation of two convolutional layers and two pooling layers,followed by three fully connected layers.As an activation function,the Re LU function was used on all layers except the output one.For the output layer,the Soft Max activation function was used.The loss function is represented by the MSE function.The SGD function was chosen as the optimizing function.Data was collected and preliminarily prepared for training the neural network.The data set for training the neural network included 5450 samples.Performance was tested on three different data sets with different cancer and healthy samples,and good results were obtained.Experiments show that the accuracy rate of training set is 90.5% – 94.3%,and the accuracy rate of testing set is 89.2 – 96.9%.The curves of testing accuracy rate and loss show that this model has been trained well.The accuracy rate for some case may reach 97.1%.With certain clinical application value,this deep learning method can be widely applied to the grading and staging of prostate cancer and other cancer tasks.The study will automate the process of diagnosing prostate cancer,increase the accuracy of determining cancerous tumors,reduce the burden on medical personnel.