The topic of this thesis is registration of image sequences captured by experimental ophthalmoscope. It contains anatomical description of the visual system as well as the description of functions of selected ophthalmoscopic devices. The next covered topic is theoretical summary of registration process, which is followed by an overview of the used methods, which forms the basis of the design and implementation of the registration algorithm in the Python programming language. After implementation, the accuracy and computational complexity of a registration was evaluated. Tests of optimalization of the proposed approach were performed with regards to the obtained results, through which sufficiently accurate registration has been achieved, evaluated on the basis of Euclidean distances, standard deviation and visual observation. In case of high-quality recorded sequences, values of Euclidean distances ranged from 0.60 to 4.07 pixels on the contrary, values higher than 20 pixels occurred in the case of poor-quality recordings. Standard deviation values in recordings with high enough resolution have not reached worse results than 4.12.

Nuclear fuel is visually inspected during regular shutdowns in order to monitor defects and long-term changes. To enable automatic comparison of images of fuel assemblies, it is crucial to perform their registration, the implementation of which has not yet been published in the scientific literature. In this work we present an analysis of image registration techniques and similarity metrics inspired by the focus operators used in autofocus and shape-from-focus. Their performance has been evaluated using a series of experiments that tested their various properties on a novel data set obtained in cooperation with the research organization Centrum výzkumu Řež. Finally, we present and discuss the results and make recommendations on which to use in which scenario.

fMRI imaging is considered key on the understanding of the brain and the mind, for this reason has been the subject of tremendous research connecting different disciplines. The intrinsic complexity of this 4-D type of data processing and analysis has been approached with every single computational perspective, lately increasing the trend to include artificial intelligence. One step critical on the fMRI pipeline is image registration. A model of Deep Networks based on Fully Convolutional Neural Networks, spatial transformation neural networks with a self-learning strategy was proposed for the implementation of a Fully deformable model image registration algorithm. Publicly available fMRI datasets with images from real-life subjects were used for training, testing and validating the model. The model performance was measured in comparison with ANTs deformable registration method with good results suggesting that Deep Learning can be used successfully for the development of the field using the basic strategy of studying the brain using the brain-self strategies.

Numerical oceanographic models are constantly improving and must be validated when improvements are made. One means of determining how to improve these models and performing validations is to compare model predictions to the future observed outcome, which is measured many ways, including satellite imagery. Comparisons of model forecasts to future satellite images result in error measurements. One common problem with modern oceanographic models is spatial error, i.e., the incorrect placement and shape of ocean features, rendering traditional error metrics such as mean-square and cross-correlation ineffective. Such problems are common in meteorological forecast verification as well, so the application of spatial error metrics have been a recently popular topic in that field of study. Spatial error metrics separate model error into a displacement component and an amplitude component, providing a more reliable assessment of numerical model inaccuracies and a more descriptive portrayal of model prediction skill.The application of spatial error metrics to oceanographic models has been sparse, and significantly further advances exist in the medical imaging and registration field. These advances are presented, along with modifications necessary for application to oceanographic model output and satellite imagery. Standard approaches and options for those methods in the literature are explored, and where the best arrangements of options are unclear, comparison studies are conducted. The first of these trials require the reproduction of synthetic displacements in conjunction with synthetic amplitude perturbations across 480 Navy Coastal Ocean Model (NCOM) temperature fields from various regions of the globe throughout 2009. Results revealed the success of certain approaches novel to both meteorology and oceanography, including B-spline transforms and mutual information. That, combined with other common methods, such as quasi-Newton optimization and land masking, could best recover the synthetic displacements under various synthetic intensity changes. The second set of trials compare temperature fields from NCOM and Navy Layered Ocean Model (NLOM), both 1/16-degree and 1/32-degree, to Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery. Lessons learned from the first trials were applied and extended. The resulting methods algorithmically reproduced portions of a previous hand-analyzed study and were successful in separating spatial from amplitude (temperature) errors.

Traditional deformable registration methods have achieved impressive performances but are computationally time-consuming since they have to optimize an objective function for each new pair of images. Very recently some learning-based approaches have been proposed to enable fast registration by learning to estimate the spatial transformation parameters directly from the input images. Here we present a method for 3D fast pairwise registration of brain MR images. We model the deformation function with B-splines and learn the optimal control points using a U-Net like CNN architecture. An inverse-consistency loss has been used to enforce diffeomorphicity of the deformation. The proposed algorithm does not require supervised information such as segmented labels but some can be used to help the registration process. We also implemented several strategies to account for the multi-resolution nature of the problem. The method has been evaluated on MICCAI 2012 brain MRI datasets, and evaluated on both similarity and invertibility of the computed transformation.

No description available.

State-of-the-art three-dimensional endo-luminal virtual colonoscopy (VC) or CT colonography (CTC) is a minimally invasive medical procedure that examines the entire colon in order to detect polyps and colorectal cancer. Most colon cancers malignantly transform from polyps, which are extra growths on the surface of the mucous membrane. Three dimensional endoscopic colon lumen interior images offered by CTC allow physicians to examine the colon interactively. Thus, CTC has several advantages over conventional optical colonoscopy including reduced risk. One of the challenging problems that prevent practical use in clinical situations is the complexity of the human colon. The colon's deformation by peristalsis and the diverse shapes of polyps make it difficult to distinguish polyps from other non-threatening entities in the colon. Hence, most CTC protocols acquire both prone and supine images to improve the visualization of the lumen wall, reduce false positives, and improve sensitivity. Comparisons between the prone and supine images can be facilitated by computerized registration between the scans. In this dissertation, two algorithms for registering colons segmented from prone and supine images are presented. First is an algorithm for three dimensional registration of the prone and supine colon when both are well distended and there is a single connected lumen. Second is another registration algorithm between colons with topological differences caused by inadequate bowel preparation or peristalsis. Such topological changes make deformable registrations of the colons difficult, and at present there are no registration algorithms which can accommodate them. The first algorithm uses feature matching of the colon centerline and a modified version of the demons deformable registration algorithm to define a deformation field between the prone and supine lumen surface. The second method utilizes embedded map representation of colon volume. The two proposed colon registration methods will contribute to improving the accuracy of the computerized registration process and increasing the versatility of the clinical use of CT colonoscopy. / Ph. D.

In the field of medical imaging, image registration methods are useful for many applications such as inter- and intra-subject morphological comparisons, creation of population atlases, delivery of precision therapies, etc. A user may want to know which is the most suitable registration algorithm that would work best for the intended application, but the vastness of medical image registration applications makes evaluation and comparison of image registration performance a non-trivial task. In general, evaluating image registration performance is not straightforward because in most image registration applications there is an absence of “Gold Standard” or ground truth correspondence map to compare against. It is therefore the primary goal of this thesis work to provide a means for recommending the most appropriate registration algorithm for a given task. One of the contributions of this thesis is to examine image registration algorithm performance at the component level. Another contribution of this thesis is to catalog the benefits and limitations of many of the most commonly used image registration evaluation approaches. One incremental contribution of this thesis was to demonstrate how existing evaluation methods can be applied in the midpoint coordinate system to evaluate some symmetric image registration algorithms such as the SyN registration algorithm. Finally, a major contribution of this thesis was to develop tools to evaluate and visualize 2D and 3D image registration shape collapse. This thesis demonstrates that many current diffeomorphic image registration algorithms suffer from the collapse problem, provides the first visualizations of the collapse problem in 3D for simple shapes and real human brain MR images, and provides the first experiments that demonstrate how adjusting image registration parameters can mitigate the collapse problem to some extent.

The digital photography is still trying to catch-up with its analogous counterpart and recording light direction is one of the most recent area of interest. The first and still the only one light-field camera for consumers, the Lytro camera, has reached market in 2011. This work introduces the light-field theory and recording with special emphasis on illustrating the principles in 2D, gives an overview of current hardware and ongoing research in the area and analyses the Lytro camera itself, describing the closed file formats and protocols it uses so that further research can be conducted. An important contribution of the work is a .NET portable library for developers, supplemented by a file editor as well as an application for wireless communication with the camera based on the library. Finally, the theory is used to discuss implications for light-field registration and linear panoramas. Powered by TCPDF (www.tcpdf.org)

Approved for public release; distribution is unlimited / Image navigation is critical to the effective use of digital imagery for meteorological and oceanographic studies. This thesis reviews various methods used to navigate imagery to the earth and investigates the accuracy of the Naval Postgraduate School (NPS) model. An explanation of how the NPS navigation process works is included for completeness. Results from 2 2 separate runs of the NPS model are studied. / http://archive.org/details/accuracyofsatell00beth / Captain, United States Marine Corps