Fully Automated Prostate Segmentation In MR Images With Combined Fuzzy C-means And U-Net Algorithms

Knowledge about the internal anatomical and volume of the prostate can increase disease diagnostic performance and improve other different and multiple clinical fields such as staging in prostate cancer and selection of appropriate treatments.Based on this fact,accurate segmentation of the prostate gland is very important computer-aided disease diagnostics.Therefore,this study presents a developed algorithm for automated prostate segmentation by utilizing MR images.To this purpose,stacks of abdominal T2-weighted MRI images are acquired.The proposed method is a U-Net convolutional neural network,which is a supervised machine learning algorithm,enhanced with fuzzy c-means(FCM)algorithm.The U-Net used in this work fundamentally consists of an encoder-decoder(i.e.downsampling upsampling)network architecture with depth of 7.To improve network performance,acquired data are augmented,whereby 1200 MR images of original data are utilized intentionally to produce 150 k MR images for training.These augmented data and labels are resized to a dimension of 256 x 256 and partitioned into 90% and 10% for training cases and validation cases,respectively.Multiple features are extracted from the input data.This method uses unsupervised feature learning of fuzzy c-means as pre-training process,and replaced the random initialization weights of traditional U-Net parameters.The method makes parameters to acquire more reasonable and significant values hence obtain higher accuracy in prostate segmentation.Based on this model,2D prostate glands are predicted.Optimal parameters for the neural networks have been established,which achieve an accuracy of 0.91 for training,and 0.84 for validation.When evaluated with 30 T2-weighted MRI images,the model obtains mean,median and standard deviation of the Dice similarity coefficient(DSC)scores of 0.85,0.88,and 0.05,respectively.The comparison is made between the proposed enhanced U-Net model and the existing U-Net model.Results show that the proposed method manages to segment prostate images precisely.The evaluation of proposed algorithm experiments is done by using the data from the Medical Image Computing and Computer-Assisted Intervention(MICCAI) Workshop.