Studies And Application Of Precision Medicine-oriented Intelligent Methods For Medical Image Registration

With the deepening and the popularity of the concept of precision medicine,and the continuous development of artificial intelligence technology and medical imaging technology,multimoda l medical images are widely used in clinical diagnosis,surgical navigation and the formulation of personalized treatment plans for patients.At the same time,with the combina tion of multimodal medical images,the medical image fusion technology is widely used to assist diagnosis.Medical image registration is the premise of image fusion.Because of the great differences between multimoda l medical images,existing multimoda l medical image registration methods mostly focused on brain,and these methods are time-consuming and difficult to be applied in clinical practice.Takeing into full account the high precision characteristics of the methods of monomodal registration,the synthetic CT was adopted into the traditional multimoda l registration.We proposed a novel framework of multimoda l registration which contains two major parts,i.e.synthetic CT generation and monomoda l registration between synthetic CT and real CT.The main achievements of research on this topic are as follows:1)Aming at the specific scene of the synthetic CT generation in the brain,we proposed a multitask maximum entropy clustering algorithm to segment MR images,and then generate high-quality synthetic CT imags.For the complex body section of abdomen and pelvis,two advanced machine learning technique i.e.transfer learning and semisupervised learning were adopted in medical image processing.We proposed a method of synthetic CT generation based on UTE-m Dixon sequence and jointly leveraging prior knowledge as well as partial supervision.Consider about that these advanced MR sequences e.g.UTE,are technically challenging,we further proposed a method of synthetic CT generation with transfer learning and semi-supervised classification jointly applied based m Dxion only.The results of experiments indicate the proposed method can generate high-quality synthetic CT in the complex body parts e.g.abdomen and pelvis.2)Inspired by the technique of synthetic CT generation from MR sequences,we proposed a MR-CT multimodal image registration method based on registration on synthetic CT.The results of experiments shows that better effect of and less time consumption of registration of the proposed method of registration than traditional methods.3)By adopting the proposed multimoda l image registration method framework in a concrete situation of the completion of MR images with unknown deformation caused by damage and pollution,we proposed a method of MR image completion guided by registration of synthetic CT.The repaired images in brain of subjects indicated that the proposed method is effectively.4)We summarize the data characteristics of MR images in the process of synthetic CT generation,and help our team to propose a method of patch learning based synthetic CT generation.In the mean time,inspired by the voxel based method of segmentation of MR images,we help our team to propose a radial basis neural network-leeraged fast traing method for the identification of organs in the regions of interesting of MR images.