Prostate Zonal Segmentation In MR Image Based On Fully Convolutional U-Net

The prostate is an accessory gland of the male reproductive system and can be divided into two parts,the central zone and the peripheral zone,from the radiologic viewpoint.Distinguishing the central zone and the peripheral zone in T2-weighted magnetic resonance images may help detect prostate cancer and is a prerequisite for quantitative analysis.Handcrafted image segmentation algorithms could not delineate the zonal boundaries accurately enough.Therefore,deep learning techniques has been used increasingly more often to crack the problem,especially the fully convolutional U-Net.In the literature,2D U-Net was often used with single slices as the analytical unit,which,however,suffered from ignoring the correlation between the neighboring slices.In the meanwhile,the use of the complete 3D sequence data for segmentation faced challenges of the high computational complexity and huge memory consumption.Therefore,the focus of this thesis is to develop a better framework that takes both advantages of the 2D U-Net with lower computational complexity and the 3D data with more complete information to segment the different zones of the prostate.In this thesis,the information between the slices is used to segment the central zone and peripheral zone of the prostate in T2-weighted magnetic resonance images,based on a fully convolutional U-Net,namely,the 2.5D ULSTM-Net.The Dice similarity coefficient(DSC)value is used to evaluate the performance of the network on a public dataset,the NCI-ISBI prostate dataset.The main contents of this paper are as follows.(1)2D U-Net prostate zonal segmentation based on single slices.After image resizing,both center cropping and online augmentation are used in model training.Fivefold cross-validation is used to choose hyper-parameters.Although timage cropping does not improve the segmentation accuracy,the memory usage is reduced substantially.The image augmentation improves the performance and the DSC values are 0.87±0.04 and 0.80±0.05,for segmentation of the center and peripheral zones,respectively.(2)3D U-Net prostate zonal segmentation based on 3D data.The 3D convolutional layer helps extract information between slices to assist segmentation.However,the 3D pooling layer in the 3D network down-samples the 3D data isometrically,causing substantial loss of information from the adjacent slices.Therefore,replacing the 3D pooling layer with the 2D pooling layer helps improve the performance.Compared with the original 3D U-Net,segmentation accuracy is significantly improved.However,due to the large memory the 3D data occupies,the model is difficult to be fully optimized,and ultimately,the performance is almost equivalent to that of the 2D U-Net.(3)2.5D ULSTM-Net prostate zonal segmentation based on neighboring slices.The convolutional layer in the decoding stage is replaced with a bidirectional Conv LSTM module,and the information in the nearest 5 adjacent slices is fused for the segmentation.Compared with the 2D and 3D networks,2.5D ULSTM-Net improves the DSC to 0.89±0.03 and 0.81±0.04,for segmentation of the center and peripheral zones,respectively.Above results demonstrate that,compared with the 2D U-Net of single slices and the modified 3D U-Net of 3D data,the 2.5D ULSTM-Net,taking consideration of the information in neighboring slices,improves the segmentation and reduces the memory requirement.The proposed method is an accurate and efficient method for prostate zonal segmentation.Tested on the NCI-ISBI public dataset,the proposed method are better than those of the state of the art methods in the literature in terms of the DSC.