Reconstruction Of Knee Joint Model Based On Deep Neural Networks And Morphological Filtering

Magnetic resonance imaging(MRI)is the main imaging method for the diagnosis of knee joint.Currently high-field MRI has obtained an increased applications in clinical diagnosis in view of its advantages of high signal-tonoise ratio(SNR)and spatial resolution.However,the local specific absorption rate(SAR)in high-field scanning is a crucial safety factor that cannot be ignored.At present,the mainstream method for estimating local SAR is to carry out electromagnetic(EM)simulation to obtain the electrical field distribution.In this situation the rapid construction of subject-specific knee model is necessary.Therefore,it is important to implement an automatic and accurate tissue segmentation in knee images.The convolutional neural network(CNN)is one type of deep neural network that has been developed rapidly at present.It has good performance and great potential in the application of image segmentation.However,the SNR of some images is relatively low and the tissue distribution and architecture of knee joint are complex,therefore it is difficult to achieve satisfied results for using a CNN network simply to segment knee images.In this paper,two segmentation methods for knee image based on U-Net are studied.Combined with morphological filtering,they are used to construct knee model separately.Therefore,more accurate estimation of local SAR can be achieved.In the first method,images of two types of modals are fed into a U-Net network for processing.Compared with using single-modal images,this method can enhance the ability of ‘memory’ of network model for some tissues in images,alleviate the influence of low SNR to some extent.In the second method,the parallel architecture of multiple U-Net networks is used to replace a single U-Net for processing,each network implements the segmentation of two tissues that have the approximate volume size.This can reduce the influence caused by the difference in tissue volume size.The research works of this paper are listed as below:(1)To eliminate holes and isolate voxels in the segmented images,a postprocessing method based on morphological filtering is implemented.This postprocessing is following the network processing for improving their output results.(2)On the basis of analyzing the characteristics of T1-and T2-weighted images of knee joint,a segmentation method based on T1-and T2-weighted images is presented.These two types of images are fed into a U-Net network simultaneously for training model parameters.The trained network is then used to process images of test set for obtaining tissue classification results.(3)In order to solve the problem caused by the unbalanced of tissue volume size and further improve the segmentation accuracy of small tissue,a method based on parallel networks is proposed.Knee tissues are divided into three groups for training according to their volume size.On this basis,images of the test set are processed.(4)Reconstruct models using the segmentation results obtained by the above two methods.The EM simulation is carried out and local SAR is calculated using the constructed models separately.In the experiments,the image segmentation performance of the presented two methods are evaluated using measurements such as DCC(Dice’s coefficient).Results show that the segmentation results of the presented methods are more close to those of manual delineation.Compared with only using a single network,the relative error of peak local SAR between our methods and manual delineation is smaller.This validates the availability of the presented methods.