Research On Registration Algorithm Of Brain Magnetic Resonance Imaging Based On U-Net

Medical image registration plays a crucial part in operation guidance,segmentation,dose accumulation,reconstruction,and registration.With the development of computer hardware and convolutional neural network(CNN),the speed of deep learning-based medical image registration is improved sharply.Since the field-of-view of CNN is limited,the CNN-based deformable image registration algorithms have not obtained satisfying performance.In addition,the folding point in registration results caused the performance of model fails to meet the clinical requirements.To address the above-mentioned issues,we take several experiments on 3D brain magnetic resonance imaging(MRI).By redesigning the feature extraction algorithm and objection function,the proposed model improved registration accuracy and generalizability.The innovations of this paper are as follows:(1)To extend the limited field-of-view of the CNN-based methods,we proposed the registration model based on attention mechanism and dubbed DAU-Net.By utilizing the multi-scale attention mechanism to extend the field-of-view,the DAU-Net obtained plentiful global feature representations.The channel attention mechanism is embedded into the corresponding layer of U-Net to improve the non-linear mapping ability of DAU-Net.In addition,we redesigned the objective function to improve the smoothness of the DVF.The experimental results demonstrate that the DAU-Net has obtained promising performance.(2)To improve the feature extraction ability and eliminate the cross-folding phenomenon that existed in the registration results,we proposed the Transformer-based registration model dubbed Trans DIR.The down-sampling module based on Transformer is designed to extract the long-range dependencies.By employing the attention up-sampling module,the Trans DIR increased the focus on region-of-interest(ROI)and suppressed the irrelevant background region.Additionally,the anti-folding penalty with Jacobin determinant is employed to eliminate the cross-folding points in registration results.The experimental results demonstrate that the evaluation metrics are improved significantly.(3)Since the noise in the feature map disturbs Transformer from capturing long-range dependencies,we proposed the registration model based on graph convolution Transformer and dubbed Graformer DIR.By utilizing the adjacent matrix of graph convolution to shield the interference of missing and spurious connections,the Graformer DIR obtained efficient longrange dependencies.The experimental results indicate that the performance of Graformer DIR is improved effectively.