Research On Human Brain MR Image Registration Method Based On U-net

Alzheimer's disease,attention deficit hyperactivity disorder and other neurological diseases deeply affect the health of every patient.These neurological disorders are slow onset and are not easily diagnosed.If they can be detected in an early stage,they can largely reduce the incidence rate of these diseases.The anatomic structure in human brain magnetic resonance(MR)images has clear pathological significance,which plays an important role in the discovery and treatment of neurological diseases.At the same time,the registration of human brain MR images is the key to the diagnosis of neurological diseases.Therefore,the study of human brain MR image registration has very important medical significance.With the continuous development of deep learning in the field of computer vision,medical image registration methods based on deep learning have achieved many impressive results,especially the one-time generation of predictive transformation parameters can be used for ultra fast registration.At the same time,in the method of one-time generation of prediction transformation parameters,a variety of encoder-decoder structure,such as U-net,is widely used.However,these models have some limitations in the process of deformation processing.The registration accuracy of registration methods based on these models needs to be improved.Based on this,this paper proposes the following three medical image registration methods based on deep learning:(1)In this paper,an unsupervised 3D medical image registration method based on nested dense connection U-net is proposed.In the process of training,this method does not need the real deformation field and other supervision information,and can predict the whole displacement vector field at one time.The proposed model uses gap filling to gradually enrich the features of skipping connection,so that the cascaded feature maps are similar,which makes the registration network easier to optimize.At the same time,the convolution of holes at the bottom of nested dense connection u-net can increase the receptive field of the model,thus increasing the maximum offset of displacement vector in each direction,which has a positive impact on the prediction of deformation field.(2)This paper proposes a 3D medical image registration method based on V-net(MSD-Vnet),which combines multi-scale skip connection,selective kernel attention mechanism and depth supervision.Specifically,the proposed MSD-Vnet uses multi-scale connection skipping to improve the positioning ability and boundary perception ability of the network for anatomical structure location information.MSD-Vnet makes each layer of encoder cascade full-scale feature mapping.Selective kernel attention mechanism 3D SK-Net adaptively adjusts the size of receptive field according to the multi-scale features of medical image,and guides the network how to select and allocate the representation of different size kernels,so as to improve the registration accuracy.At the same time,depth supervision can help the network learn better and prevent the gradient disappearing.(3)This paper proposes a V-net-based V-shaped multi-path network(VV-net)for3D medical image registration.The registration model can be trained end-to-end by stacking V-nets.Specifically,the moving image is distorted by two V-Nets in turn,and the additional V-Net is used to provide supplementary information for the first two V-Nets to form a V-shaped network,so that the moving image can be better aligned with the fixed image.At the same time,the depth supervision auxiliary branch is added to the proposed model to prevent over fitting.The accuracy of registration is improved by using the progressive registration and information supplement.The proposed three medical image registration methods are verified on the common human brain MR image datasets,ADNI,ADNI,and the datasets composed of four datasets,ADNI,ABIDE,ADHD-200 and OASIS,respectively.The first two medical image registration methods are evaluated on the common human brain MR image data set ADNI,and the third method is verified on the ADNI data set,ABIDE data set,ADHD-200 data set and OASIS data set,and the measure is the dice similarity coefficient(DSC).Experimental results show that,compared with the current popular methods,the proposed method achieves satisfactory results.