Skip to Main content Skip to Navigation

Recalage d'images déformées en groupe pour l'estimation du mouvement en imagerie médicale 4D

Abstract : This doctoral thesis develops methods to estimate patient's motion, voluntary and involuntary (organs' motion), in order to correct for motion in spatiotemporal tomographic medical images. As an experimental paradigm we consider the problem of motion estimation in Diffusion-Weighted Magnetic Resonance Imaging (DWI), an imaging modality sensitive to the diffusion of water molecules in the body. DWI is used for the evaluation of lymphoma patients, since water diffuses differently in healthy tissues and in lesions. The effect of water diffusion can be better depicted through a parametric map, the so-called apparent diffusion coefficient (ADC map), created based on a series of DW images of the same patient (3D image sequence), acquired in time during scanning. Such a parametric map has the potentiality to become an imaging biomarker in DWI and provide physicians with complementary information to current state-of-the-art FDG-PET imaging reflecting quantitatively glycose metaboslism. Using the ADC as an imaging biomarker is appealing due to the fact that it is derived non-invasively, it doesn't require any exogenous contrast agents, it doesn't use ionizing radiation, it is quantitative, it can be obtained relatively rapidly, and it is easily incorporated into routine patient evaluations. Retention though of the spatial information derived by the ADC map requires the correction for motion, which is achieved through co-registering the DW images of the sequence. Our contributions are three fold. First, we propose a group-wise deformable image registration method especially designed for motion correction in DWI, as it is guided by a physiological model describing the diffusion process taking place during image acquisition. Our method derives an ADC map of higher accuracy in terms of depicting the gradient of the water molecules' diffusion in comparison to the corresponding map derived by common practice or by other model-free group-wise image registration methods. Second, we show that by imposing spatial constraints on the computation of the ADC map, the tumours in the image can be even better characterized in terms of classifying them into the different types of the disease. Third, we show that a correlation between DWI and FDG-PET should exist by examining the correlation between statistical features extracted by the smooth ADC map derived by our deformable registration method, and recommendation scores on the malignancy of the lesions, given by experts based on an evaluation of the corresponding FDG-PET images of the patient. In more detail, the first part of the thesis introduces the problem of motion correction in 4D medical imaging (4DMI), while it focuses on the corresponding problem in DWI. Moreover, in that introductory chapter, the notion of deformable medical image registration is as well introduced, followed by a statement on the goals and the roadmap of this doctoral thesis. Furthermore, the second part of the thesis is devoted on a literature review on proposed methods coping with motion estimation in 4DMI. As this particular registration problem involves a sequence of images to be co-registered, it is referred to as group-wise image registration problem, which is more challenging than the common pair-wise registration problem in terms of balancing computational efficiency and solution accuracy. Then in the third part of the thesis, our first group-wise deformable image registration method is introduced and described, which incorporates the temporal dimension (reflecting the change of signal amplitude in time) of the acquisition process. Temporal consistency on the physiological model, as well as deformation smoothness, is imposed through a pair-wise Markov Random Field (MRF) formulation, towards producing anatomically meaningful representations of the 3D images. The performance of the proposed method is compared, using three different validation criteria, to two different model-free group-wise registration approaches; one that penalizes the absolute differences in the intensities and one that penalizes the intensity range among the images on corresponding regions. A dataset consisting of twenty-five (25) patients, each scanned with 3 "b-values", was used to evaluate the method's accuracy. The proposed registration method outperformed the other two registration approaches, making it a very promising method for highlighting the importance of ADC as an imaging biomarker. In the fourth part of the thesis, a novel joint parameter estimation method is introduced and described, which consists of the main contribution of this doctoral thesis. It is based as well on a pair-wise Markov Random Field (MRF) formulation that jointly registers the DW images and models the spatiotemporal diffusion. Spatial smoothness on the ADC map, as well as spatiotemporal deformation smoothness, is imposed in that framework too. In contrast to our first proposed method, the ADC map in this formulation is computed explicitly and not implicitly as it was the case before through fitting the physiological model. A dataset consisting of thirty-eight (38) patients, this time each scanned with 5 "b-values", was used to evaluate the method's accuracy, as well as the effect of using more than three "b-values" to create the ADC map. To this scope, simulated DWI data were created in order to optimize the parameters of the method. The registration performance is compared to our first proposed methods and a state-of-the art registration approach in terms of obtained fitting error of the diffusion model in the core of the tumor. Our results reveal a marginally better performance of both our methods when compared against the standard ADC map used in clinical practice, a state-of-the-art pair-wise registration method, as well as the model-free group-wise registration methods, which indicates their potential as means for extracting imaging biomarkers. The main qualities of our frameworks lie in their computational efficiency and versatility. The discrete nature of the formulations renders the frameworks modular in terms of iconic similarity measures. Finally, in the fifth and last part of the thesis, we examined whether we can extract potential imaging biomarkers (image-related features) by the ADC maps derived by our proposed registration methods. The evaluation is based on two different statistical analysis, one in which we examine whether we could automatically classify the lesions into the different types of the disease and another one where we examined the correlation between the extracted features and the recommendation grading scores derived by FDG-PET. Automatic disease-type classification (Hodgkin and Non-Hodgkin) based on imaging biomarkers extracted using our registration method achieved a classification accuracy of approximately 73.3%, yielding a 7.5% increase in respect to the standard approach (biomarkers from unregistered DW images). Moreover, significant correlation between diffusion-based features and PET-based staging scores was revealed both after two cycles of chemotherapy and in the end of treatment (corr=0.52 and corr=0.78 respectively, p-value < 0.01 in both) by the proposed method that was not observed without registration. Applying linear regression and using the features showing the best correlation in the afore-mentioned study, staging scores for the patients were predicted that had a significant correlation with the corresponding PET-based ones (corr=0.71 and corr=0.88, respectively, p-value $<$ 0.01 in both), while the corresponding coefficients of determination were 0.68 and 0.91, meaning that the model explains 68% and 91% of the variability in the PET-based staging scores respectively. No accurate prediction was attained without registration.
Document type :
Complete list of metadata
Contributor : Evgenios N. Kornaropoulos <>
Submitted on : Sunday, August 27, 2017 - 7:59:47 PM
Last modification on : Saturday, May 1, 2021 - 3:49:13 AM


  • HAL Id : tel-01577683, version 1


Evgenios N. Kornaropoulos. Recalage d'images déformées en groupe pour l'estimation du mouvement en imagerie médicale 4D. Bio-informatique [q-bio.QM]. Ecole Centrale Paris, 2017. Français. ⟨tel-01577683⟩



Record views


Files downloads