Advanced Kernel Methods For Remote Sensing Image Classification

Résumé Suite aux recentes avancées technologiques, les archives d'images digitales ont connu une croissance qualitative et quantitative sans précédent. Malgré les énormes possibilités qu'elles offrent, ces avancées posent de nouvelles questions quant au traitement des masses de données saisies. Cette question est à la base de cette Thèse: les problèmes de traitement d'information digitale à très haute résolution spatiale et/ou spectrale y sont considérés en recourant à des approches d'apprentissage statistique, les méthodes à noyau. Cette Thèse étudie des problèmes de classification d'images, c'est à dire de catégorisation de pixels en un nombre réduit de classes refletant les propriétés spectrales et contextuelles des objets qu'elles représentent. L'accent est mis sur l'efficience des algorithmes, ainsi que sur leur simplicité, de manière à augmenter leur potentiel d'implementation pour les utilisateurs. De plus, le défi de cette Thèse est de rester proche des problèmes concrets des utilisateurs d'images satellite sans pour autant perdre de vue l'intéret des méthodes proposées pour le milieu du machine learning dont elles sont issues. En ce sens, ce travail joue la carte de la transdisciplinarité en maintenant un lien fort entre les deux sciences dans tous les développements proposés. Quatre modèles sont proposés: le premier répond au problème de la haute dimensionalité et de la redondance des données par un modèle optimisant les performances en classification en s'adaptant aux particularités de l'image. Ceci est rendu possible par un système de ranking des variables (les bandes) qui est optimisé en même temps que le modèle de base: ce faisant, seules les variables importantes pour résoudre le problème sont utilisées par le classifieur. Le manque d'information étiquétée et l'incertitude quant à sa pertinence pour le problème sont à la source des deux modèles suivants, basés respectivement sur l'apprentissage actif et les méthodes semi-supervisées: le premier permet d'améliorer la qualité d'un ensemble d'entraînement par interaction directe entre l'utilisateur et la machine, alors que le deuxième utilise les pixels non étiquetés pour améliorer la description des données disponibles et la robustesse du modèle. Enfin, le dernier modèle proposé considère la question plus théorique de la structure entre les outputs: l'intègration de cette source d'information, jusqu'à présent jamais considérée en télédétection, ouvre des nouveaux défis de recherche. Advanced kernel methods for remote sensing image classification Devis Tuia Institut de Géomatique et d'Analyse du Risque September 2009 Abstract The technical developments in recent years have brought the quantity and quality of digital information to an unprecedented level, as enormous archives of satellite images are available to the users. However, even if these advances open more and more possibilities in the use of digital imagery, they also rise several problems of storage and treatment. The latter is considered in this Thesis: the processing of very high spatial and spectral resolution images is treated with approaches based on data-driven algorithms relying on kernel methods. In particular, the problem of image classification, i.e. the categorization of the image's pixels into a reduced number of classes reflecting spectral and contextual properties, is studied through the different models presented. The accent is put on algorithmic efficiency and the simplicity of the approaches proposed, to avoid too complex models that would not be used by users. The major challenge of the Thesis is to remain close to concrete remote sensing problems, without losing the methodological interest from the machine learning viewpoint: in this sense, this work aims at building a bridge between the machine learning and remote sensing communities and all the models proposed have been developed keeping in mind the need for such a synergy. Four models are proposed: first, an adaptive model learning the relevant image features has been proposed to solve the problem of high dimensionality and collinearity of the image features. This model provides automatically an accurate classifier and a ranking of the relevance of the single features. The scarcity and unreliability of labeled. information were the common root of the second and third models proposed: when confronted to such problems, the user can either construct the labeled set iteratively by direct interaction with the machine or use the unlabeled data to increase robustness and quality of the description of data. Both solutions have been explored resulting into two methodological contributions, based respectively on active learning and semisupervised learning. Finally, the more theoretical issue of structured outputs has been considered in the last model, which, by integrating outputs similarity into a model, opens new challenges and opportunities for remote sensing image processing.

Bilateral Sequential NAION: A Comparison of Visual Outcomes in Fellow Eyes Using Quantitative Analysis of Goldmann Visual Fields

Purpose: Previous studies of the visual outcome in bilateral non-arteritic anterior ischemic optic neuropathy (NAION) have yielded conflicting results, specifically regarding congruity between fellow eyes. Prior studies have used measures of acuity and computerized perimetry but none has compared Goldmann visual field outcomes between fellow eyes. In order to better define the concordance of visual loss in this condition, we reviewed our cases of bilateral sequential NAION, including measures of visual acuity, pupillary function and both pattern and severity of visual field loss.Methods: We performed a retrospective chart review of 102 patients with a diagnosis of bilateral sequential NAION. Of the 102 patients, 86 were included in the study for analysis of final visual outcome between the affected eyes. Visual function was assessed using visual acuity, Goldmann visual fields, color vision and RAPD. A quantitative total visual field score and score per quadrant was analyzed for each eye using the numerical Goldmann visual field scoring method previously described by Esterman and colleagues. Based upon these scores, we calculated the total deviation and pattern deviation between fellow eyes and between eyes of different patients. Statistical significance was determined using nonparametric tests.Results: A statistically significant correlation was found between fellow eyes for multiple parameters, including logMAR visual acuity (P = 0.0101), global visual field (P = 0.0001), superior visual field (P = 0.0001), and inferior visual field (P = 0.0001). In addition, the mean deviation of both total (P = 0.0000000007) and pattern (P = 0.000000004) deviation analyses was significantly less between fellow eyes ("intra"-eyes) than between eyes of different patients ("inter"-eyes).Conclusions: Visual function between fellow eyes showed a fair to moderate correlation that was statistically significant. The pattern of vision loss was also more similar in fellow eyes than between eyes of different patients. These results may help allow better prediction of visual outcome for the second eye in patients with NAION. These findings may also be useful for evaluating efficacy of therapeutic interventions.

Performance of a new gadolinium-based intravascular contrast agent in free-breathing inversion-recovery 3D coronary MRA.

In three-dimensional (3D) coronary magnetic resonance angiography (MRA), the in-flow contrast between the coronary blood and the surrounding myocardium is attenuated as compared to thin-slab two-dimensional (2D) techniques. The application of a gadolinium (Gd)-based intravascular contrast agent may provide an additional source of signal and contrast by reducing T(1blood) and supporting the visualization of more distal or branching segments of the coronary arterial tree. In six healthy adults, the left coronary artery (LCA) system was imaged pre- and postcontrast with a 0.075-mmol/kg bodyweight dose of the intravascular contrast agent B-22956. For imaging, an optimized free-breathing, navigator-gated and -corrected 3D inversion recovery (IR) sequence was used. For comparison, state-of-the-art baseline 3D coronary MRA with T(2) preparation for non-exogenous contrast enhancement was acquired. The combination of IR 3D coronary MRA, sophisticated navigator technology, and B-22956 allowed for an extensive visualization of the LCA system. Postcontrast, a significant increase in both the signal-to-noise ratio (SNR; 46%, P < 0.05) and contrast-to-noise ratio (CNR; 160%, P < 0.01) was observed, while vessel sharpness of the left anterior descending (LAD) artery and the left coronary circumflex (LCX) were improved by 20% (P < 0.05) and 18% (P < 0.05), respectively.

Relaxivity of Gd-based contrast agents on X nuclei with long intrinsic relaxation times in aqueous solutions.

The relaxivity of commercially available gadolinium (Gd)-based contrast agents was studied for X-nuclei resonances with long intrinsic relaxation times ranging from 6 s to several hundred seconds. Omniscan in pure 13C formic acid had a relaxivity of 2.9 mM(-1) s(-1), whereas its relaxivity on glutamate C1 and C5 in aqueous solution was approximately 0.5 mM(-1) s(-1). Both relaxivities allow the preparation of solutions with a predetermined short T1 and suggest that in vitro substantial sensitivity gains in their measurement can be achieved. 6Li has a long intrinsic relaxation time, on the order of several minutes, which was strongly affected by the contrast agents. Relaxivity ranged from approximately 0.1 mM(-1) s(-1) for Omniscan to 0.3 for Magnevist, whereas the relaxivity of Gd-DOTP was at 11 mM(-1) s(-1), which is two orders of magnitude higher. Overall, these experiments suggest that the presence of 0.1- to 10-microM contrast agents should be detectable, provided sufficient sensitivity is available, such as that afforded by hyperpolarization, recently introduced to in vivo imaging.

MRI-based prostate brachytherapy seed localization.

A magnetic resonance imaging (MRI) pulse sequence and a corresponding image processing algorithm to localize prostate brachytherapy seeds during or after therapy are presented. Inversion-Recovery with ON-resonant water suppression (IRON) is an MRI methodology that generates positive contrast in regions of magnetic field susceptibility, as created by prostate brachytherapy seeds. Phantoms comprising of several materials found in brachytherapy seeds were created to assess the usability of the IRON pulse sequence for imaging seeds. Resulting images show that seed materials are clearly visible with high contrast using IRON, agreeing with theoretical predictions. A seed localization algorithm to process IRON images demonstrates the potential of this imaging technique for seed localization and dosimetry.

Changes in volume, surface estimate, three-dimensional shape and total number of neurons of the human primary visual cortex from midgestation until old age.

Macroscopic features such as volume, surface estimate, thickness and caudorostral length of the human primary visual cortex (Brodman's area 17) of 46 human brains between midgestation and 93 years were studied by means of camera lucida drawings from serial frontal sections. Individual values were best fitted by a logistic function from midgestation to adulthood and by a regression line between adulthood and old age. Allometric functions were calculated to study developmental relationships between all the features. The three-dimensional shape of area 17 was also reconstructed from the serial sections in 15 cases and correlated with the sequence of morphological events. The sulcal pattern of area 17 begins to develop around 21 weeks of gestation but remains rather simple until birth, while it becomes more convoluted, particularly in the caudal part, during the postnatal period. Until birth, a large increase in cortical thickness (about 83% of its mean adult value) and caudorostral length (69%) produces a moderate increase in cortical volume (31%) and surface estimate (40%) of area 17. After birth, the cortical volume and surface undergo their maximum growth rate, in spite of a rather small increase in cortical thickness and caudorostral length. This is due to the development of the pattern of gyrification within and around the calcarine fissure. All macroscopic features have reached the mean adult value by the end of the first postnatal year. With aging, the only features to undergo significant regression are the cortical surface estimate and the caudorostral length. The total number of neurons in area 17 shows great interindividual variability at all ages. No decrease in the postnatal period or in aging could be demonstrated.

MBIS: multivariate Bayesian image segmentation tool.

We present MBIS (Multivariate Bayesian Image Segmentation tool), a clustering tool based on the mixture of multivariate normal distributions model. MBIS supports multichannel bias field correction based on a B-spline model. A second methodological novelty is the inclusion of graph-cuts optimization for the stationary anisotropic hidden Markov random field model. Along with MBIS, we release an evaluation framework that contains three different experiments on multi-site data. We first validate the accuracy of segmentation and the estimated bias field for each channel. MBIS outperforms a widely used segmentation tool in a cross-comparison evaluation. The second experiment demonstrates the robustness of results on atlas-free segmentation of two image sets from scan-rescan protocols on 21 healthy subjects. Multivariate segmentation is more replicable than the monospectral counterpart on T1-weighted images. Finally, we provide a third experiment to illustrate how MBIS can be used in a large-scale study of tissue volume change with increasing age in 584 healthy subjects. This last result is meaningful as multivariate segmentation performs robustly without the need for prior knowledge.

An active contour-based atlas registration model applied to automatic subthalamic nucleus targeting on MRI: method and validation.

This paper presents a new non parametric atlas registration framework, derived from the optical flow model and the active contour theory, applied to automatic subthalamic nucleus (STN) targeting in deep brain stimulation (DBS) surgery. In a previous work, we demonstrated that the STN position can be predicted based on the position of surrounding visible structures, namely the lateral and third ventricles. A STN targeting process can thus be obtained by registering these structures of interest between a brain atlas and the patient image. Here we aim to improve the results of the state of the art targeting methods and at the same time to reduce the computational time. Our simultaneous segmentation and registration model shows mean STN localization errors statistically similar to the most performing registration algorithms tested so far and to the targeting expert's variability. Moreover, the computational time of our registration method is much lower, which is a worthwhile improvement from a clinical point of view.

Active Contour-Based Segmentation of Head and Neck with Adaptive Atlas Selection

This paper presents automated segmentation of structuresin the Head and Neck (H\&N) region, using an activecontour-based joint registration and segmentation model.A new atlas selection strategy is also used. Segmentationis performed based on the dense deformation fieldcomputed from the registration of selected structures inthe atlas image that have distinct boundaries, onto thepatient's image. This approach results in robustsegmentation of the structures of interest, even in thepresence of tumors, or anatomical differences between theatlas and the patient image. For each patient, an atlasimage is selected from the available atlas-database,based on the similarity metric value, computed afterperforming an affine registration between each image inthe atlas-database and the patient's image. Unlike manyof the previous approaches in the literature, thesimilarity metric is not computed over the entire imageregion; rather, it is computed only in the regions ofsoft tissue structures to be segmented. Qualitative andquantitative evaluation of the results is presented.

Segmentation of Head and Neck Lymph Node Regions for Radiotherapy Planning Using Active Contour-Based Atlas Registration

In this paper, we present the segmentation of the headand neck lymph node regions using a new active contourbased atlas registration model. We propose to segment thelymph node regions without directly including them in theatlas registration process; instead, they are segmentedusing the dense deformation field computed from theregistration of the atlas structures with distinctboundaries. This approach results in robust and accuratesegmentation of the lymph node regions even in thepresence of significant anatomical variations between theatlas-image and the patient's image to be segmented. Wealso present a quantitative evaluation of lymph noderegions segmentation using various statistical as well asgeometrical metrics: sensitivity, specificity, dicesimilarity coefficient and Hausdorff distance. Acomparison of the proposed method with two other state ofthe art methods is presented. The robustness of theproposed method to the atlas selection, in segmenting thelymph node regions, is also evaluated.

Statistical assessment of dataset shift and model portability in multi-angle in-track image acquisitions

In this study we propose an evaluation of the angular effects altering the spectral response of the land-cover over multi-angle remote sensing image acquisitions. The shift in the statistical distribution of the pixels observed in an in-track sequence of WorldView-2 images is analyzed by means of a kernel-based measure of distance between probability distributions. Afterwards, the portability of supervised classifiers across the sequence is investigated by looking at the evolution of the classification accuracy with respect to the changing observation angle. In this context, the efficiency of various physically and statistically based preprocessing methods in obtaining angle-invariant data spaces is compared and possible synergies are discussed.

An integer-based score to predict functional outcome in acute ischemic stroke: The ASTRAL score.

OBJECTIVE: To develop and validate a simple, integer-based score to predict functional outcome in acute ischemic stroke (AIS) using variables readily available after emergency room admission. METHODS: Logistic regression was performed in the derivation cohort of previously independent patients with AIS (Acute Stroke Registry and Analysis of Lausanne [ASTRAL]) to identify predictors of unfavorable outcome (3-month modified Rankin Scale score >2). An integer-based point-scoring system for each covariate of the fitted multivariate model was generated by their β-coefficients; the overall score was calculated as the sum of the weighted scores. The model was validated internally using a 2-fold cross-validation technique and externally in 2 independent cohorts (Athens and Vienna Stroke Registries). RESULTS: Age (A), severity of stroke (S) measured by admission NIH Stroke Scale score, stroke onset to admission time (T), range of visual fields (R), acute glucose (A), and level of consciousness (L) were identified as independent predictors of unfavorable outcome in 1,645 patients in ASTRAL. Their β-coefficients were multiplied by 4 and rounded to the closest integer to generate the score. The area under the receiver operating characteristic curve (AUC) of the score in the ASTRAL cohort was 0.850. The score was well calibrated in the derivation (p = 0.43) and validation cohorts (0.22 [Athens, n = 1,659] and 0.49 [Vienna, n = 653]). AUCs were 0.937 (Athens), 0.771 (Vienna), and 0.902 (when pooled). An ASTRAL score of 31 indicates a 50% likelihood of unfavorable outcome. CONCLUSIONS: The ASTRAL score is a simple integer-based score to predict functional outcome using 6 readily available items at hospital admission. It performed well in double external validation and may be a useful tool for clinical practice and stroke research.

Inhibition in early Alzheimer's disease: an fMRI-based study of effective connectivity.

Changes of functional connectivity in prodromal and early Alzheimer's disease can arise from compensatory and/or pathological processes. We hypothesized that i) there is impairment of effective inhibition associated with early Alzheimer's disease that may lead to ii) a paradoxical increase of functional connectivity. To this end we analyzed effective connectivity in 14 patients and 16 matched controls using dynamic causal modeling of functional MRI time series recorded during a visual inter-hemispheric integration task. By contrasting co-linear with non co-linear bilateral gratings, we estimated inhibitory top-down effects within the visual areas. The anatomical areas constituting the functional network of interest were identified with categorical functional MRI contrasts (Stimuli>Baseline and Co-linear gratings>Non co-linear gratings), which implicated V1 and V3v in both hemispheres. A model with reciprocal excitatory intrinsic connections linking these four regions and modulatory inhibitory effects exerted by V3v on V1 optimally explained the functional MRI time series in both subject groups. However, Alzheimer's disease was associated with significantly weakened intrinsic and modulatory connections. Top-down inhibitory effects, previously detected as relative deactivations of V1 in young adults, were observed neither in our aged controls nor in patients. We conclude that effective inhibition weakens with age and more so in early Alzheimer's disease.

A forensic science perspective on the role of images in crime investigation and reconstruction

This article presents a global vision of images in forensic science. The proliferation of perspectives on the use of images throughout criminal investigations and the increasing demand for research on this topic seem to demand a forensic science-based analysis. In this study, the definitions of and concepts related to material traces are revisited and applied to images, and a structured approach is used to persuade the scientific community to extend and improve the use of images as traces in criminal investigations. Current research efforts focus on technical issues and evidence assessment. This article provides a sound foundation for rationalising and explaining the processes involved in the production of clues from trace images. For example, the mechanisms through which these visual traces become clues of presence or action are described. An extensive literature review of forensic image analysis emphasises the existing guidelines and knowledge available for answering investigative questions (who, what, where, when and how). However, complementary developments are still necessary to demystify many aspects of image analysis in forensic science, including how to review and select images or use them to reconstruct an event or assist intelligence efforts. The hypothetico-deductive reasoning pathway used to discover unknown elements of an event or crime can also help scientists understand the underlying processes involved in their decision making. An analysis of a single image in an investigative or probative context is used to demonstrate the highly informative potential of images as traces and/or clues. Research efforts should be directed toward formalising the extraction and combination of clues from images. An appropriate methodology is key to expanding the use of images in forensic science.