Research On Non-rigid Registration For Lung DCE-MRI

Lung cancer is the most common cause of cancer death worldwide.Early detection of lung cancer is vital for reducing mortality.Dynamic Contrast Enhanced Magnetic Resonance Imaging(DCE-MRI)is a promising non-invasive technology,which can visualize the anatomical morphology and provide physiological function information including blood flow.This technology is of great significance for the diagnosis of benign and malignant tumors.The acquisition of DCE-MRI images generally takes several minutes to monitor contrast agent uptake and washout kinetics.Unfortunately,during the acquisition,patient motion(e.g.breathing,heartbeat and pulsation)may cause misalignment between images acquired at different times,which is defined as motion artifact.If the motion artifacts appear on the lesion,it may affect the diagnosis accuracy.Considering the deformation types of patient motion are mainly about elastic deformation,non-rigid image registration can be used to compensate for motion artifacts between DCE-MRI images in the field of medical image processing.However,developing a specific non-rigid registration method for DCE-MRI is challenging due to the changes in image intensity caused by contrast agent.The local intensity changes often lead to unrealistic deformations in the traditional non-rigid registration method.To provide more accurate registration results for DCE-MRI and avoid unrealistic deformation in registration,the some key issues in DCE-MRI image registration were deeply studied in the thesis based on the lung DCE-MRI images.The main content and contributions of this thesis are summarized as follows:(1)Non-rigid image registration is prone to unrealistic deformations when the local image contrast changes fast between DCE-MRI images.To address this problem,a novel landmark-based image registration method was proposed,which used automatically detected landmark information for reducing unrealistic deformations in registration.In order to improve the landmark matching accuracy on DCEMRI images,a multi-scale local rigid matching(MsLRM)was proposed based on the assumption that a global non-rigid deformation has the characteristics of local rigidity.In addition,a multi-scale matching strategy was introduced in MsLRM considering the deformation degrees varied for different images.More specifically,the proposed MsLRM was performed from large scale to small scale achieved by setting different sampling intervals.After that,a two-stage method is designed for outlier removal before the landmark pairs are incorporated into an intensity-based registration method considering the registration is highly sensitive to outliers.Experimental results showed that the proposed landmark-based registration showed an improved registration performance on lung DCE-MRI images compared with the existing landmark-based registration methods.(2)To improve the robustness of non-rigid registration to the local contrast changes,a landmark-based group-wise registration was proposed by combining adaptive weighting landmark constraints and a robust principal component analysis(RPC A).To reduce the influence of the contrast changes,RPC A was iteratively used to separate motion from contrast changes.Landmark pairs were detected on the resulting low-rank images and then incorporated into an intensity-based registration through a constraint term.Then,the motion was reduced by applying the resulting deformation to original DCE-MRI images.To reduce the negative impact of inaccurate landmark pairs on registration accuracy,an adaptive landmark weight calculation was proposed with the assumption of local motion consistency.Landmark pairs with high matching accuracy were assigned higher weights to increase their impact on registration.Experimental results showed that the proposed method was robust to the intensity changes caused by a contrast agent and had an improved registration accuracy on lung DCE-MRI images,especially for the images before maximum contrast enhancement.(3)Due to the imperfect separation between motion and contrast changes,the data decomposition-based registration tended to produce implausible deformations and had the drawback of high computational complexity.To address this problem,a novel registration method based on image-to-image(I2I)translation was proposed.The problem of separation of motion and contrast changes was formulated as an I2I translation problem.I2I was used to generate simulated DCE-MRI images without contrast changes,and the resulting deformations can be directly derived from the generated images using FFD registration.To deal with multiple image domains,a multi-domain image-to-image translation(MDIT)with multi branch structure was designed for outputting diverse images for all available image domains.To increase the style diversity and avoid mode collapse,a style mapping network in MDIT was designed for generating random style from Gaussian noise.To deal with the problem of large differences in intensity values between different domains,the L1 loss was normalized to balance the training weight between different domains.Experimental results showed improvements of our registration framework against the existing group-wise registration in terms of both accuracy and efficiency.