Development of image registration methods for aided diagnosis and monitoring in computed tomography

• Κύριος συγγραφέας: Βλαχόπουλος, Γεώργιος
• Άλλοι συγγραφείς: Κωσταρίδου, Ελένη
• Μορφή: Thesis
• Γλώσσα: Greek
• Έκδοση: 2017
• Θέματα: Computed tomography; interstitial lung disease; Image registration; Υπολογιστική τομογραφία; Διάμεση πνευμονική νόσος; Αντιστοίχηση εικόνων; 616.075 722

The term interstitial lung disease (ILD) refers to more than 200 chronic lung disorders that are classified to the same group because of their similar clinical, radiological, physiologic and pathologic features [1]. Accurate chest CT quantification of ILD extent is crucial for patient management, since no robust biomarkers for monitoring disease progression exist [2]. Chest computed tomography, (CT) imaging is a powerful modality for detection, diagnosis, and follow up of interstitial lung Disease. CT bases assessment of ILD extent estimation is performed by several semi-quantitative scoring systems, estimating extent, utilizing high-resolution CT (HRCT) protocols [3], which are deprived of information regarding localization of disease, demonstrating moderate inter- and intra-observer agreement [4]. Thus, ILD extent estimation (quantification) and progression by means of advanced image analysis methods, is emerging offering increased accuracy and reproducibility. [5][6][7]. In the frame of quantitative image-based follow-up ILD monitoring and response to therapy in CT, image registration methods have the important role to ensure that any measured volume change between follow-up scans is caused by ILD patterns change and not by patient’s breathing or positioning during CT scanning. Multiresolution non-rigid registration, capable for capturing local lung tissue deformations, accounts for a commonly used approach for lung CT registration utilized in all resolution levels (fully non rigid schemes), or in combination with rigid transforms applied in low resolution levels (hybrid schemes). Up-to-now only one study has reported on ILD follow-up monitoring, utilizing deformable image registration (DIR) [8]. However, registration scheme parameters were adopted rather than obtained from a systematic analysis taking in to account their effect to registration accuracy. Evaluation methods for DIR accuracy assessment in CT of the thorax has receive considerable attention in the frame of adaptive radiotherapy [9], however remains an open issue [10][11]. The current thesis addresses, selection of optimal registration schemes in case of ILD follow-up in CT. The current thesis addresses selection of suitable multiresolution registration schemes for CT based follow up analysis of ILD with the following specific objectives: • Development of an evaluation methodology to obtain high accuracy schemes • Analyzing the effect of scheme components on both registration accuracy and time efficiency • Investigation of the effect of mass preserving cost function to registration accuracy and time efficiency. • Investigation of the effect of regularization term in artificially generated follow up data. The evaluation methodology, considers two stages: the first stage utilizes artificially warped ILD follow-up data to identify candidate registration schemes, while in the second stage the performance of candidate schemes, identified in the first stage, are verified with clinical follow-up data . At the first stage, the determinant of the Jacobian matrix of the each voxel of deformation field is the main tool to identify schemes including folding areas to be excluded from subsequent analysis. The performance of the remaining schemes is assessed and ranked in terms of their displacement error in two anatomical regions, i.e. Normal Lung Parenchyma (NLP) and ILD affect regions, by means of Euclidean distance of homologus points between baseline and register follow up pair. Statistical analysis was performed to select near optimal schemes, considering both NLP and ILD ranking lists of candidate schemes. In the second stage, evaluation of registration accuracy of candidate registration schemes is verified on their clinical follow up volumetric scans using the displacement error as well. Finally selected registration schemes, was also performed by statistical analysis, considering both NLP and ILD ranking lists of near optimal schemes. A clinical dataset consisting of 10 pairs of CT scans corresponding to 10 patients diagnosed with ILD secondary to connective tissue diseases, radiologically manifested with ground glass and reticular patterns, at two different times, abstaining in time approximately two years was acquired. The total extent of ILD presence in these 10 scans ranges from 5% to 80% (mean value: 32%). Artificial warped data [12] is introduced to ensure that registration error is caused by registration algorithm alone and not by intrinsic data variability. Artificially warped data were generated using a single level non rigid thin plate kernel spline model and applied to baseline lung segments, resulting in artificial follow up lung segment simulating realistically lung deformations, preserving size and slice thickness of the original data set. The basic components of a typical registration schemes are the transformation model, the cost function the optimizer and the type of pyramid used. A total of 128 registration schemes was generated, by considering the following combinations of components: Four (4) transformations: Euler Transform - ET, Similarity Transform - SM, Affine Transform - AT, and 3rd order B-spline Transform - BST, applied to the first 3 resolution levels of the pyramid in order to obtain a coarse initial alignment, while the 3rd order B-Spline transform was utilized for the 4th resolution level, corresponding to the highest image resolution, in all schemes generated. Although fully non-rigid transformation models seem the natural choice for lung field registration due to the elastic nature of the lung tissue, hybrid schemes (including rigid and non rigid transformation models in different resolution levels), are also been recently proposed. As the optimizer accounts as for a critical component of the registration process, 4 gradient decent optimizers were utilized for the optimization step: SGD, regular step gradient decent - RSGD, adaptive stochastic gradient decent - ASGD and finite difference gradient decent - FDGD. Additionally, two 2 types of pyramids were considered: Gaussian Pyramid - GP, that applies smoothing and downsampling by a factor of 2 in all three dimensions, and Recursive Pyramid – RP, that applies no smoothing but only downsampling by a factor of 2 in all three dimensions. Finally, 4 different cost functions were considered (Sum of Square Difference – SSD, Normalized Correlation Coefficient – NCC, Mutual Information – MI, and Normalized Mutual Information – NMI). Taking in to account, the susceptibility of SSD cost function to air quantity inside the lungs [13][14], due to varying breathing phase, a mass preserving variant of SSD was also considered in a separate experiment . All registration schemes evaluated in this study, are intensity based and utilize the multiresolution approach to avoid local minimum traps and speed-up calculations, using Elastix version 4.5, based on the open source software Insight Toolkit (ITK) version 4.0. Additionally, the impact of regularization constrains to select candidate schemes is addressed in a final experiment, considering schemes excluded in the first stage of the evaluation methodology. [15] Specifically, a bending energy regularization term is added to each one of the above 4 cost functions was applied to artificially warped data, to constrain irregular deformations (folding areas). Taking into account the variability of the hardware used, due to large scale of the experiment, an empirical scale of computational time performance was introduced. Taking into account that the major parameters affecting registration time performance are the selected pyramid and the type of optimizer used, 4 scales of time performance are considered from the slowest (1) to the fastest (4). 16 out of 128 near optimal registration schemes registration were initially obtained by the first first evaluation stage based on artificially generated follow up data. These schemes obtained sub-millimeter registration accuracies in terms of average distance errors 0.18 ± 0.01 mm for NLP and 0.20 ± 0.01 mm for ILD, respectively. Verification of registration accuracy in terms of average distance error in clinical follow-up data was in the range of 1.985-2.156 mm and 1.966-2.234 mm, for NLP and ILD affected regions respectively, excluding schemes with statistically significant lower performance (Wilcoxon signed-ranks test, p<0.05), resulting in 13 finally selected registration schemes, highlighting the efficiency of the stage of the evaluation stage. The observed difference in registration accuracy between the two evaluation stages may be attributed to the fact that the artificially warped scan pairs simulate lung deformations as a result of breathing, without taking into account disease progression/regression. All 13 finally selected registration schemes in case of ILD CT follow-up analysis are hybrid and include the ASGD optimizer [16], providing high accuracy and time efficiency. Additionally, the proposed pyramid is RP, as it was included in 10 out of 13 finally selected registration schemes. Regarding the impact of mass preserving cost function to registration accuracy, statistical analysis of the obtained results indicates that adding a mass preserving term in the SSD cost function, has no practical contribution in registration performance, in case of full inspiration breath-hold protocol in the framework of ILD CT follow up analysis, while its omission significantly reduces the computational cost of about a factor of 3. The introduction of a regularization term, contributes to 6 additional near optimal registration schemes in the first stage, however without statistically significantly higher accuracy compared to the 16 original selected near optimal schemes. Finally, considering computational costs, the introduction of binding energy penalty term, increases computational cost by a factor of 4 providing however about 4 times higher computational cost.