Research And Application Of Multi-modal 3D Medical Image Alignment Algorithm

Kidney cancer is a highly prevalent urological disease that seriously threatens human health.For patients with kidney cancer,surgical resection can achieve radical cure while preserving part of the kidney unit and leaving the kidney function in a normal state.Therefore,it is especially important to assess the residual kidney function after partial removal of the kidney.The location of the tumor can be localized in the enhanced CT images,while SPECT images can provide functional information of the kidney,and the relationship between these two modalities is to be established by the plain CT images.In this paper,we focus on calculating the residual kidney function by aligning the abdominal kidney with the plain CT images and the enhanced CT images,and finally mapping the tumor information to the SPECT images.The research directions in this paper are divided into two categories,one is to solve the problem of layer-to-layer non-correspondence due to different spatial locations of medical images in different modalities,and the other is to solve the problem of how to better improve the alignment performance when the medical image annotation dataset is not perfect.(1)For the different spatial positions of plain scan CT images and enhanced CT images taken at different times,resulting in the problem that layers do not correspond to layers,this paper proposes a registration method based on cubic spline interpolation.The method is divided into two steps,because the ultimate goal of this study is to locate the tumor position on the plain CT image,so the enhanced CT sequence image is selected for interpolation,in addition,because the layer thickness of the enhanced CT image is thinner than the plain CT image,the interpolation between the enhanced CT tomography images can be more approximate to the actual cross-section.Based on the above reasons,this paper interpolates the enhanced CT sequence image,and then performs affine registration with the non-contrast CT sequence image.(2)To address the problem of imperfect medical image annotation dataset,this paper proposes a multimodal medical image alignment method based on unsupervised learning.In this paper,we use the voxelmorph model as the backbone network,add a 3D-UNet segmentation network to construct semantic segmentation loss,back-propagate it into the backbone network,guide its training,and continuously optimize it to further strengthen the feature representation capability.And a loss function suitable for multimodal medical image alignment is designed to improve the alignment performance.Through the experimental analysis of the alignment of enhanced CT images and flat-scan CT images of kidney,the average Dice coefficient is improved by 2.3%,which verifies the effectiveness and accuracy of the method.(3)In the process of diagnosis and treatment of renal tumors,there is a lack of automatic computer system to determine tumor position on non-contrast CT images,and this paper designs a multimodal medical image registration system to automatically judge tumor position in non-contrast CT images.Based on the design of QT and VTK frameworks,the system uses the cubic spline interpolation registration algorithm proposed in this paper and the registration algorithm based on unsupervised learning to make automatic judgment.The system is divided into data import module,data preprocessing module,coarse registration module and fine registration module,and each module is combined with each other to jointly complete the judgment of tumor position in the non-contrast CT image.Through the above method,this paper completes the research on the method of multimodal 3D medical image alignment,designs a multimodal 3D medical image intelligent alignment system,conducts experimental analysis of each work,proves the effectiveness of the proposed method,and completes the judgment of tumor location in plain CT images.