| In medical image segmentation,alignment,imaging histology and computer-aided diagnosis,high resolution medical images can significantly improve experimental accuracy.Due to the limitations of MRI equipment and image acquisition time,it is usually not possible to acquire 3D MRI images of high quality.Most of these acquired images have high resolution within MRI slices and poor resolution between slices,which seriously affects the experimental accuracy of MRI image analysis tasks.The study of 3D MRI super-resolution reconstruction is therefore of great importance to the development and advancement of modern medical MRI image analysis.For mono-axial head MRI image data,this paper proposes a multi-scale densely connected convolution-based MRI super-resolution reconstruction method.Under the condition of layer spacing of 2 mm,the input data size is 5×256×256 3D medical images and the output is 9×256×256 3D medical images,and the new images are reconstructed between adjacent layers to complete the super-resolution reconstruction between slices.The model uses densely connected convolution as the base module for multi-scale information fusion,enhances feature propagation in the network through dense connections,reduces the resolution of the image through pooling to obtain image information at different scales,uses three down-sampling,and three transposition convolutions,and uses the dense convolution module before each down-sampling and after the transposition convolution,and connects images of the same size in the process of down-sampling and transposition convolution.The images were jump-jointed during the down-sampling and transposition convolution.The peak signal-to-noise ratio of the reconstructed images at a layer spacing of 2 mm was 43.86 d B and the structural similarity was 0.995.For multi-axial head MRI image data,this paper proposes a multi-axial MRI image preprocessing method to spatially align MRI images of multiple axes.According to whether spatial position relationships are recorded during MRI imaging,a spatial alignment method based on spatial geometry calculation and a global residual network MRI image alignment model based on channel attention are proposed,respectively.The input of the alignment model is one MRI image in each of the three axes,and the output of the model is the transverse and longitudinal coordinates of the intersection points of the three axial MRI images in their respective planes.The model uses multiple channel attention cascades and a global residual structure,and the experimental results show that the alignment error of different axial MRI data is 1.11-2.20 mm.For multi-axial head MRI image data,based on the spatial alignment,this paper proposes an MRI super-resolution model based on a hybrid framework of Transformer and CNN.The input data of the model are spatially aligned coronal,sagittal and cross-sectional slices covered with MRI cavity post-nodal data,and the output is cavity internal data.The model introduces a self-attention mechanism to focus on the image texture protection during super-resolution reconstruction,which effectively protects the texture structure of the reconstructed image and focuses the super-resolution problem on the smaller 3D cavity interior,facilitating the application of the model.This paper compares the reconstruction effect under various layer spacing,and in order to enhance the training effect of the network this paper proposes a migration learning method.The experimental results show that the peak signal-to-noise ratio of the reconstructed images at a layer spacing of 4 mm is 36.81 d B and the structural similarity is 0.887.In this paper,two spatial alignment methods and two super-resolution reconstruction methods are proposed for 3D MRI images with large layer spacing between slices to achieve3 D super-resolution reconstruction of single-axis and multi-axis MRI images with various layer spacing conditions,and the reconstructed images are clearer and expand the data set for medical image analysis tasks. |
PDF Full Text Request