The Study Of Medical Image Non-linear Registration Based On Machine Learning

Non-linear medical image registration can well establish the anatomical correspondences across images,and it is a fundamental processing step in many medical image analysis tasks.Non-linear image registration plays an important role in many clinical applications,such as population analysis,longitudinal studies,diseases diagnosis,disease therapy,and etc.The aim of non-linear registration is to establish the spatial transformation through deformation field,which can build the corresponding relationship across two images.Thus,non-linear deformable registration is a complicated nonlinear optimization problem.With the increasing image demention and data scale,the medical imaging data are more complex and diverse.This makes the conventional non-linear registration methods less accurate,robust and applicable when dealing with diverse data.To better solve the non-linear registration problem on multi-modal data,longitudinal data and the data with large anatomical variations,based on several real clinical requirements,this thesis proposes several learning-based non-linear deformable registration algorithms,in which different machine learning techniques are well applied to tackle the challenging registration problems.The main works of this thesis are summarized as follows.First,a learning-based multi-modal non-linear registration algorithm is proposed,to solve the pelvic CT and MRI registration problem in prostate cancer radiation therapy.There are two challenges for pelvic CT and MRI registration: the large appearance gap between CT and MRI and the large anatomical variations on main pelvic organs.To address these issues,a bidirectional image synthesis based multi-modal non-linear registration method is proposed to effectively register the pelvic CT and MRI images.This algorithm can well facilitate the accurate prostate cancer radiation therapy in real clinical application.The details of this work are described as follows.Based on conventional random forest learning algorithm,a structured random forest andmulti-target random forest methods are proposed to establish the bi-directional imagesynthesis model: not only synthesizing CT from MRI(MRI→CT),but also synthesizingMRI from CT(CT→MRI).The image synthesis can bridge the appearance gap betweenCT and MRI.During image synthesis,the bi-directional synthesis manner can wellpreserve the whole anatomical information in both modalities,which provides an unbiasedreference when performing multi-modal image registration.Based on bi-directional image synthesis,a dual-core steered multi-modal non-linearregistration algorithm is proposed.By using bi-directional image synthesis,the complexmulti-modal registration can be simplified to a mono-modal registration problem.Next,inorder to guarantee the registration accuracy,a dual-core deformation fusion algorithm isproposed,which can sufficiently use the complementary anatomical information in bothCT and MRI.The local deformation can be effectively estimated by iterating the dual-coredeformation fusion algorithm through CT modality core and MRI modality core,andfinally improve the accuracy of CT and MRI registration.Based on bi-directional image synthesis,a region-adaptive multi-modal non-linearregistration algorithm is proposed.In order to effectively use the salient information ofbone region in CT and soft tissue in MRI,this thesis proposes a region-adaptive non-linearregistration algorithm.In this algorithm,the CT modality is used to guide the registrationof bone region and the MRI modality is used to guide the registration of soft tissue.In thisway,during the registration,the salient,but complementary anatomical details of the twomodalities can be jointly used,to accurately estimate the local deformation of all the pelvicorgans.Second,a longitudinal infant brain non-linear registration algorithm is proposed based on brain development model.The accurate longitudinal infant brain registration is a basic processing step for infant brain analysis and brain-related diseases diagnosis.Due to the fast brain development and white matter myelination,the appearance and anatomical structures of infant brain images are changed dynamically within the first year of life.This thesis proposes a multi-task random forest learning algorithm,which is used to establish two models to capture the infant brain development: the brain structure development model and the appearance evolution model.These two models are built from the images acquired at 2 weeks,3 months,6 months and 9 months after birth to the images acquired at 12 months,respectively.By using the improved multi-task random forest learning method,the learning models can be established more accurately,and meanwhile effectively preserve the infant brain topology.Based on the brain structure development model and the appearance evolution model,the large appearance and anatomical variations between images at different time points can be effectively compensated,where the non-linear registration task on infant images can be simplified.Finally,the longitudinal infant brain images can be accurately registered.Third,a deep learning-based brain image non-linear registration algorithm is proposed,which is important to the analysis of diverse brain-related diseases and studies.Conventional non-linear registration algorithms need sufficiently iterative optimization,along with careful parameter tuning.When the images have large anatomical variations,the registration accuracy may decline.To address these issues,this thesis proposes a cue-aware deep regression network,where a convolutional neural network is used to establish a non-linear registration model.Based on this model,the deformation field can be effectively estimated by inputting the to-beregistered image pair.In order to enhance the awareness of the network for the special registration task,an auxiliary contextual cue is generated via the proposed data-driven network,and this cue can provide informative reference when training the registration network.Additionally,in order to improve the accuracy and generalization of the non-linear registration model,a key-points guided balanced sampling strategy is further proposed.This sampling strategy can generate a more completed and well-functioning training set,which can help avoid the overfitting issues.The trained model can be directly applied to predict the deformation field of a pair of images,without the need of iterative optimization and parameter tuning.When dealing with the images that have large anatomical variations,this algorithm can consistently work well.Compared with the conventional registration algorithms,the proposed non-linear registration algorithm can significantly improve the registration accuracy and robustness,which makes it more applicable in diverse clinical applications.Based on different clinical backgrounds,i.e.,prostate cancer radiation therapy,infant brain analysis and adult brain-related analysis,this thesis proposes several learning-based non-linear registration algorithms,in order to fulfill the diverse clinical requirements.Based on random forest and convolutional neural network learning algorithms,this thesis proposes several improved learning algorithms,which are well applied to help solve different non-linear registration problems,such as multi-modal non-linear registration,longitudinal non-linear image registration and registration for images with large anatomical variations.The study of this thesis is of great importance to make the modern medical image analysis more precise and intelligent.