Tesis sobre el tema "Automated Segmentation Method"

Les hyperintensités de la substance blanche (HSB) sont de plus en plus prises en compte dans le suivi clinique des personnes âgées et/ou des patients atteints de démences, et sont cruciales chez les patients atteints de Sclérose en Plaques (SEP). Des méthodes d'analyse ont été proposées pour aider à quantifier ces lésions à grande échelle, afin de mieux comprendre les mécanismes sous-jacents de ces pathologies. Cependant, à notre connaissance, il n'y a pas de consensus aujourd'hui sur la méthode à utiliser et aucune méthode n'est validée sur ces deux types de sujets. Cette thèse présente plusieurs outils et leurs validations dans le but de mieux caractériser les HSB. Tout d'abord, WHASA-3D (Tran et al., 2022) est une nouvelle méthode de segmentation automatique des HSB adaptée pour les données 3D T2-FLAIR et aux patients SEP dans un cadre multicentrique. C'est une amélioration majeure de WHASA (Samaille et al. 2012). Les performances de WHASA-3D sont comparées ici avec six méthodes de la littérature avec leurs paramètres par défaut et optimisés lorsque cela est possible. Deux extensions ont alors été développées dans le but d'apporter une aide lors du diagnostic et du suivi clinique des patients. WHASA-Spatial est une extension pour caractériser spatialement les HSB fournies par WHASA-3D selon quatre classes (périventriculaire, infratentorielle, juxtacorticales/corticales, profondes). La classification a été évaluée visuellement sur 104 sujets SEP et a montré des résultats très satisfaisants. Enfin, WHASA-Longitudinal, est une extension permettant de segmenter automatiquement les nouvelles lésions ou lésions élargies entre deux acquisitions successives. Les performances de cette méthode se sont révélées satisfaisantes au niveau de la volumétrie et une piste est proposée pour améliorer le comptage de lésions. Ces résultats doivent être confirmés sur une plus grande base de sujets
White matter hyperintensities (WMH) are more and more taken into account in the clinical monitoring of elderly subjects and/or dementia patients, and are crucial in patients with Multiple Sclerosis (MS). Automated methods have been proposed to better quantify these lesions on a large scale, in order to better understand the underlying mechanisms of these pathologies. However, to our knowledge, no automated method has reached consensus today for the segmentation of WMH, and no method has been validated on these two types of subjects. This thesis introduces several tools and their validations in order to better characterize WMH. First of all, WHASA-3D (Tran et al. 2022) is a new automated method for WMH segmentation adapted for 3D T2-FLAIR data and MS patients in a multicenter setting. It is a major improvement of WHASA (Samaille et al. 2012). WHASA-3D's performances are here compared with six state-of-the-art methods with their default parameters and optimized settings, when possible. Two extensions have then been developped to support the clinician for patient diagnosis and clinical monitoring. WHASA-Spatial is an extension for the automatic spatial characterization of WMH provided by WHASA-3D according to four classes (periventricular, infratentorial, juxtacortical/cortical, deep). The visual assessment on 104 MS subjects showed that the global classification was very satisfactory. Finally, WHASA-Longitudinal, is an extension that allows the automatic segmentation of new or enlarged lesions between two successive acquisitions. The performance of this method was satisfactory for volume agreement and a solution is proposed and needs to be investigated to improve new lesion count. These results need to be confirmed on a larger number of subjects

Breast cancer is the second leading cause of death of women worldwide. Accurate lesion boundary detection is important for breast cancer diagnosis. Since many crucial features for discriminating benign and malignant lesions are based on the contour, shape, and texture of the lesion, an accurate segmentation method is essential for a successful diagnosis. Ultrasound is an effective screening tool and primarily useful for differentiating benign and malignant lesions. However, due to inherent speckle noise and low contrast of breast ultrasound imaging, automatic lesion segmentation is still a challenging task. This research focuses on developing a novel, effective, and fully automatic lesion segmentation method for breast ultrasound images. By incorporating empirical domain knowledge of breast structure, a region of interest is generated. Then, a novel enhancement algorithm (using a novel phase feature) and a newly developed neutrosophic clustering method are developed to detect the precise lesion boundary. Neutrosophy is a recently introduced branch of philosophy that deals with paradoxes, contradictions, antitheses, and antinomies. When neutrosophy is used to segment images with vague boundaries, its unique ability to deal with uncertainty is brought to bear. In this work, we apply neutrosophy to breast ultrasound image segmentation and propose a new clustering method named neutrosophic l-means. We compare the proposed method with traditional fuzzy c-means clustering and three other well-developed segmentation methods for breast ultrasound images, using the same database. Both accuracy and time complexity are analyzed. The proposed method achieves the best accuracy (TP rate is 94.36%, FP rate is 8.08%, and similarity rate is 87.39%) with a fairly rapid processing speed (about 20 seconds). Sensitivity analysis shows the robustness of the proposed method as well. Cases with multiple-lesions and severe shadowing effect (shadow areas having similar intensity values of the lesion and tightly connected with the lesion) are not included in this study.

We propose two different methods for image segmentation with the objective of marking contaminated regions in images from biochemical tests. The contaminated regions consists of thin hair or fibers and the purpose of this thesis is to eliminate the tedious task of masking the contaminated regions by hand by implementing automatic hair masking. Initially an algorithm based on Morphological Image Processing is presented, followed by solving the problem of pixelwise classification using a Convolutional Neural Network (CNN). Finally, the performance of each implementation is measured by comparing the segmented images with labelled images which are considered to be the ground truth. The result shows that both implementations have strong potential at successfully performing semantic segmentation on the images from the biochemical tests.

Dans cette thèse, nous abordons deux problèmes principaux, à savoir l'évaluation quantitative des algorithmes de segmentation de maillages ainsi que la segmentation de maillages par apprentissage en exploitant le facteur humain. Nous proposons les contributions suivantes : - Un benchmark dédié à l'évaluation des algorithmes de segmentation de maillages 3D. Le benchmark inclut un corpus de segmentations vérités-terrains réalisées par des volontaires ainsi qu'une nouvelle métrique de similarité pertinente qui quantifie la cohérence entre ces segmentations vérités-terrains et celles produites automatique- ment par un algorithme donné sur les mêmes modèles. De plus, nous menons un ensemble d'expérimentations, y compris une expérimentation subjective, pour respectivement démontrer et valider la pertinence de notre benchmark. - Un algorithme de segmentation par apprentissage. Pour cela, l'apprentissage d'une fonction d'arête frontière est effectué, en utilisant plusieurs critères géométriques, à partir d'un ensemble de segmentations vérités-terrains. Cette fonction est ensuite utilisée, à travers une chaîne de traitement, pour segmenter un nouveau maillage 3D. Nous montrons, à travers une série d'expérimentations s'appuyant sur différents benchmarks, les excellentes performances de notre algorithme par rapport à ceux de l'état de l'art. Nous présentons également une application de notre algorithme de segmentation pour l'extraction de squelettes cinématiques pour les maillages 3D dynamiques.

A fundamental automatic target recognition (ATR) system can be composed of an object segmentation stage, followed by feature extraction from those objects produced by segmentation, and finally classification of these object features. The capability of such a system in terms of classification success is therefore limited not only by the quality of the feature extraction and classification methods used, but also by the quality of the initial object segmentation. In this thesis, a novel architecture is described which uses two stages of segmentation. This allows image features derived after a primary segmentation stage to influence the parameters of a secondary segmentation stage which is applied to the same image area. This is aimed at allowing improved, and locally optimised, segmentation of those objects which were poorly segmented by the primary segmentation stage. To enable the implementation of the system, a probability density estimate function is used as a method of detecting novelty in objects presented for classification. This is found to be a non-ideal solution, although useful in the context of the application concerned. The development of all the system components, and ultimately the full ATR system, is described with experimental results derived from real-world infrared imagery. From this work, conclusions are drawn as to the usefulness of a such a two-stage segmentation architecture; specifically, the clutter rejection flexibility and the potential ability for the system to locally optimise segmentation on a per object basis are highlighted.

This thesis aims to contribute to the area of visual tracking, which is the process of identifying an object of interest through a sequence of successive images. The thesis explores kernel-based statistical methods, which map the data to a higher dimensional space. A pre-image framework is provided to find the mapping from the embedding space to the input space for several manifold learning and dimensional learning algorithms. Two algorithms are developed for visual tracking that are robust to noise and occlusions. In the first algorithm, a kernel PCA-based eigenspace representation is used. The de-noising and clustering capabilities of the kernel PCA procedure lead to a robust algorithm. This framework is extended to incorporate the background information in an energy based formulation, which is minimized using graph cut and to track multiple objects using a single learned model. In the second method, a robust density comparison framework is developed that is applied to visual tracking, where an object is tracked by minimizing the distance between a model distribution and given candidate distributions. The superior performance of kernel-based algorithms comes at a price of increased storage and computational requirements. A novel method is developed that takes advantage of the universal approximation capabilities of generalized radial basis function neural networks to reduce the computational and storage requirements for kernel-based methods.

The objective of the proposed research is to develop methods that couple an expert user's guidance with automatic image segmentation and registration algorithms. Often, complex processes such as fire, anatomical changes/variations in human bodies, or unpredictable human behavior produce the target images; in these cases, creating a model that precisely describes the process is not feasible. A common solution is to make simplifying assumptions when performing detection, segmentation, or registration tasks automatically. However, when these assumptions are not satisfied, the results are unsatisfactory. Hence, removing these, often times stringent, assumptions at the cost of minimal user input is considered an acceptable trade-off. Three milestones towards reaching this goal have been achieved. First, an interactive image segmentation approach was created in which the user is coupled in a closed-loop control system with a level set segmentation algorithm. The user's expert knowledge is combined with the speed of automatic segmentation. Second, a stochastic point set registration algorithm is presented. The point sets can be derived from simple user input (e.g. a thresholding operation), and time consuming correspondence labeling is not required. Furthermore, common smoothness assumptions on the non-rigid deformation field are removed. Third, a stochastic image registration algorithm is designed to capture large misalignments. For future research, several improvements of the registration are proposed, and an iterative, landmark based segmentation approach, which couples the segmentation and registration, is envisioned.

Thesis (Ph. D.)--Electrical and Computer Engineering, Georgia Institute of Technology, 2008.
Committee Chair: Allen Tannenbaum; Committee Member: Anthony Yezzi; Committee Member: Marc Niethammer; Committee Member: Patricio Vela; Committee Member: Yucel Altunbasak.

Segmentation of scanned tissue volumes of three-dimensional (3D) computed tomography (CT) images often involves—at least partially—some manual process, as there is no standardized automatic method. There is a need to develop fully automatic approaches, not only to improve the objectivity of the task, but also to increase the overall speed of the segmentation process. Here we extend a 3D local binary patterns (LBP) based trabecular bone segmentation method with adaptive local thresholding and additional segmentation parameters to make it more robust yet still perform adequately when compared to traditional user-assisted segmentation. We estimate parameters for the new automated adaptive multiscale LBP-based 3D segmentation method (AMLM) in our experimental setting, and have two micro-CT (μCT) scanned bovine trabecular bone tissue volumes segmented by both the AMLM and two experienced users. Comparison of the results shows superior performance of the AMLM suggesting the strong potential for this solution to perform automatic bone segmentation
Skannattujen kudosrakenteiden segmentointi kolmiulotteisista (3D) tomografiakuvista tehdään usein ainakin osittain manuaalisesti, sillä standardoitua automaattista menetelmää ei ole. Täysin automatisoitujen lähestymistapojen kehitys on tarpeen, sillä se parantaisi sekä segmentoinnin objektiivisuutta että sen kokonaisnopeutta. Tässä työssä laajennamme automatisoitua local binary patterns (LBP) -perustaista trabekulaarisen luun 3D-segmentointimenetelmää adaptiivisella paikallisella kynnystyksellä ja segmentoinnin lisäparametreilla tavoitteenamme vahvistaa menetelmää mutta säilyttää silti riittävä suorituskyky verrattuna perinteiseen käyttäjäavusteiseen segmentointiin. Arvioimme koejärjestelyssämme parametrit uudelle automatisoidulle adaptiiviselle moniasteikkoiselle LBP-pohjaiselle 3Dsegmentointimenetelmälle (AMLM), ja teetämme sekä AMLM:n avulla että kahden kokeneen käyttäjän toimesta binäärisegmentoinnit kahdelle mikrotietokonetomografialla (μTT) tuotetulle kuvalle naudan trabekulaarisesta luukudoksesta. Tulosten vertailu osoittaa AMLM:n suorituskyvyltään selkeästi paremmaksi, mikä antaa vahvan viitteen tämän menetelmän soveltuvuudesta automatisoituun luusegmentointiin

Modern medical imaging techniques have revolutionized health care over the last decades, providing clinicians with high-resolution 3D images of the inside of the patient's body without the need for invasive procedures. Detailed images of the vascular anatomy can be captured by angiography, providing a valuable source of information when deciding whether a vascular intervention is needed, for planning treatment, and for analyzing the success of therapy. However, increasing level of detail in the images, together with a wide availability of imaging devices, lead to an urgent need for automated techniques for image segmentation and analysis in order to assist the clinicians in performing a fast and accurate examination. To reduce the need for user interaction and increase the speed of vascular segmentation,  we propose a fast and fully automatic vascular skeleton extraction algorithm. This algorithm first analyzes the volume's intensity histogram in order to automatically adapt the internal parameters to each patient and then it produces an approximate skeleton of the patient's vasculature. The skeleton can serve as a seed region for subsequent surface extraction algorithms. Further improvements of the skeleton extraction algorithm include the expansion to detect the skeleton of diseased arteries and the design of a convolutional neural network classifier that reduces false positive detections of vascular cross-sections. In addition to the complete skeleton extraction algorithm, the thesis presents a segmentation algorithm based on modified onion-kernel region growing. It initiates the growing from the previously extracted skeleton and provides a rapid binary segmentation of tubular structures. To provide the possibility of extracting precise measurements from this segmentation we introduce a method for obtaining a segmentation with subpixel precision out of the binary segmentation and the original image. This method is especially suited for thin and elongated structures, such as vessels, since it does not shrink the long protrusions. The method supports both 2D and 3D image data. The methods were validated on real computed tomography datasets and are primarily intended for applications in vascular segmentation, however, they are robust enough to work with other anatomical tree structures after adequate parameter adjustment, which was demonstrated on an airway-tree segmentation.

Sphere-forming assay is an in-vitro technique to assess the self-renewal and differentiation potential of a homogenous or heterogenous population of cells. This technique is commonly used in the stem cell and cancer biology fields to assess the ability of a cell that is capable of self-proliferation and differentiation. (Schmitt, 2011, Lombaert et al., 2008) To detect proliferative growth, Ki-67, a marker of proliferation, is used in immunofluorescence staining of sphere-forming cells. The current gold standard methodology to quantify cell proliferation is to manually count the cells on images obtained using confocal microscopy. However, the reproducibility, the inter- and intra-subject variability, and the time requirement for manually counting cells are often major challenges for researchers. In this study, we propose a semi-automated cell segmentation algorithm using the FARSIGHT toolbox, to automatically count the individual three-dimensional (3-D) cell nuclei. The present work focused on the investigation of two aspects of the algorithm performance: sensitivity and specificity. We grouped images by sphere size to test specificity of the algorithm. For the sensitivity analysis, we tested the segmentation algorithm on both raw uncalibrated images and calibrated images using Fiji ImageJ software. We found that the proposed algorithm could efficiently identify cells and cell boundaries to overcome the background noise. Finally, statistical analysis showed the differentiation index had low percentage matching between the proposed method and the manual counting method.

In this thesis, we have focused on the image pre-processing in order to enhance the image information. The main aspects of this enhancement rely on removing any image noise and correcting any intensity bias induced by the scanner. Besides, we also contributed with a new technique based on a multispectral, adaptive, region growing algorithm in order to segment the brain from the rest of the head. We include, as a pre-processing step, the image registration process, in which we proposed a novel pipeline by using information from multiple modalities to improve the results of this process. Furthermore, we have also studied the current techniques for the detection and segmentation of WML, proposing a new method based on a previous proposal. Therefore, we presented a tool able to automatically detect and segment WML of Multiple sclerosis and Lupus patients.
En aquesta tesi ens centrem, per una part, en el pre-processat de la imatge per tal d'eliminar el soroll i corregir les inhomogeneïtats en les intensitats, ambdós errors introduïts per l'escàner. A més hem contribuït també amb una nova tècnica basada en un algoritme de “región growing” per tal de segmentar el cervell de dins de tota la imatge del cap. Incloem com a pre-processat el registre d'imatges, on hem proposat una “pipeline" mitjançant la informació de múltiples modalitats per tal de millorar els resultats d'aquest procés. Per altra banda, hem estudiat també les tècniques actuals de detecció i segmentació de lesions en la matèria blanca, proposant un mètode nou basat en anteriors propostes. Així doncs, presentem una eina automàtica capaç de detectar i segmentar lesions en la matèria blanca de pacients d'Esclerosi Múltiple i Lupus.

Context: The main purpose of this thesis is to build an automatic handwritten digit recognition method for the recognition of connected handwritten digit strings. To accomplish the recognition task, first, the digits were segmented into individual digits. Then, a digit recognition module is employed to classify each segmented digit completing the handwritten digit string recognition task. In this study, different machine learning methods, which are SVM, ANN and CNN architectures are used to achieve high performance on the digit string recognition problem. In these methods, images of digit strings are trained with the SVM, ANN and CNN model with HOG feature vectors and Deep learning methods structure by sliding a fixed size window through the images labeling each sub-image as a part of a digit or not. After the completion of the segmentation, to achieve the complete recognition of handwritten digits.Objective: The main purpose of this thesis is to find out the recognition performance of the methods. In order to analyze the performance of the methods, data is needed to be used for training using machine learning methods. Then digit data is tested on the desired machine learning technique. In this thesis, the following methods are performed: Implementation of HOG Feature extraction method with SVM Implementation of HOG Feature extraction method with ANN Implementation of Deep Learning methods with CNN Methods: This research will be carried out using two methods. The first research method is the ¨Literature Review¨ and the second ¨Experiment¨. Initially, a literature review is conducted to get a clear knowledge on the algorithms and techniques which will be used to answer the first research question i.e., to know which type of data is required for the machine learning methods and the data analysis is performed. Later on, with the knowledge of RQ1, Experimentation is conducted to answer the RQ2, RQ3, RQ4. Quantitative data is used to perform the experimentation because qualitative data which obtains from case-study and survey cannot be used for this experiment method as it contains non-numerical data. In this research, an experiment is conducted to find the best suitable machine learning method from the existing methods. As mentioned above in the objectives, an experiment is conducted using SVM, ANN, and CNN. By considering the results obtained from the experiment, a comparison is made on the metrics considered which results in CNN as the best method suitable for Documents Images. Results: Compare the results for SVM, ANN with HOG Feature extraction and the CNN method by using segmented results. Based on the Experiment results it is found that SVM and ANN have some drawbacks like low accuracy and low performance in the recognition of documented images. So, the other method i.e., CNN has greater performance with high accuracy. The following are the results of the recognition rates of each method. SVM performance - 39% ANN performance - 37% CNN performance - 71%. Conclusion: This research concentrates on providing an efficient method for recognition of automatic handwritten digits recognition. Here a sample training data is treated with existing machine learning and deep learning methods like SVM, ANN, and CNN. By the results obtained from the experimentation, it clearly is shown that the CNN method is much efficient with 71% performance when compared to ANN and SVM methods. Keywords: Handwritten Digit Recognition, Handwritten Digit Segmentation, Handwritten Digit Classification, Machine Learning Methods, Deep Learning, Image processing on document images, Support Vector Machine, Conventional Neural Networks, Artificial Neural Networks

Aquesta tesi s'emmarca dins del pla integral de prevenció precoç de la Retinopatia Diabètica (RD) posat en marxa pel govern espanyol seguint les recomanacions de l'Organització Mundial de la Salut de promoure iniciatives que consciencien sobre la importància de fer revisions oculars regulars entre les persones amb diabetis. Per tal de poder determinar el nivell de retinopatia diabètica cal localitzar i identificar diferents tipus de lesions a la retina. Per poder fer-ho, cal que primer s'eliminin de la imatge les estructures anatòmiques normals de l'ull (vasos sanguinis, disc òptic i fòvea) a fi de fer més visibles les anomalies. Aquesta tesi s'ha centrat en aquest pas de neteja de la imatge. En primer lloc, aquesta tesi proposa un nou marc per a la segmentació ràpida i automàtica del disc òptic basat en la Teoria del Portafoli de Markowitz. En base a aquesta teoria es proposa un model innovador de fusió de colors capaç d'admetre qualsevol metodologia de segmentació en el camp de la imatge mèdica. Aquest enfoc s'estructura com una etapa de pre-processament potent i en temps real que es podria integrar-se a la pràctica clínica diària, permetent accelerar el diagnòstic de la DR a causa de la seva simplicitat, rendiment i velocitat. La segona contribució d'aquesta tesi és un mètode per fer simultàniament una segmentació dels vasos sanguinis i la detecció de la zona avascular foveal, reduint considerablement el temps de processament d'imatges. A més a més, el primer component de l'espai de color xyY (que representa els valors de crominància) és el que predomina en l'estudi dels diferents components de color desenvolupat en aquesta tesi, centrat en la segmentació dels vasos sanguinis i la detecció de fòvea. Finalment, es proposa una recopilació automàtica de mostres per fer una interpolació estadística del color i que són utilitzades en l'algorisme de segmentació de Convexity Shape Prior. La tesi també proposa un altre mètode de segmentació dels vasos sanguinis que es basa en una selecció de característiques efectiva basada arbres de decisions. S'ha aconseguit trobar les 5 característiques més rellevants per segmentar aquestes estructures oculars. La validació mitjançant tres tècniques de classificació diferents (arbres de decisions, xarxes neuronals i màquines de suport vectorial).
Esta tesis se enmarca dentro del plan integral de prevención contra la Retinopatía Diabética (RD), ejecutado por el Gobierno de España alineado a las políticas de la Organización Mundial de la Salud para promover iniciativas que conciencien a la población con diabetes sobre la importancia de exámenes oculares de manera periódica. Para poder determinar el nivel de retinopatía diabética hace falta localizar e identificar diferentes tipos de lesiones en la retina. Para conseguirlo primero se han de eliminar de la imagen las estructures anatómicas normales del ojo (vasos sanguíneos, disco óptico y fóvea) para hacer visibles las anomalías. Esta tesis se ha centrado en este paso de limpieza de la imagen. En primer lugar, esta tesis propone un novedoso enfoque para la segmentación rápida y automática del disco óptico basado en la Teoría de Portafolio de Markowitz. En base a esta teoría se propone un innovador modelo de fusión de color capaz de soportar cualquier metodología de segmentación en el campo de las imágenes médicas. Este enfoque se estructura como una etapa de preprocesamiento potente y en tiempo real que podría integrarse en la práctica clínica diaria para acelerar el diagnóstico de RD debido a su simplicidad, rendimiento y velocidad. La segunda contribución de esta tesis es un método para segmentar simultáneamente los vasos sanguíneos y detectar la zona avascular foveal, reduciendo considerablemente el tiempo de procesamiento para tal tarea. Adicionalmente, la primera componente del espacio de color xyY (que representa los valores de crominancia) es la que predomina del estudio de las diferentes componentes de color realizado en esta tesis para la segmentación de vasos sanguíneos y la detección de la fóvea. Finalmente, se propone una recolección automática de muestras para interpolarlas basadas en la información estadística de color y que a su vez son la base del algoritmo Convexity Shape Prior. La tesis también propone otro método de segmentación de vasos sanguíneos basado en una selección efectiva de características soportada en árboles de decisión. Se ha conseguido encontrar las 5 características más relevantes para la segmentación de estas estructuras oculares. La validación utilizando tres técnicas de clasificación (árbol de decisión, red neuronal artificial y máquina de soporte vectorial).
This thesis is framed within the comprehensive plan for early prevention of Diabetic Retinopathy (DR) launched by the Spain government following the World Health Organization to promote initiatives that raise awareness of the importance of regular eye exams among people with diabetes. To determine the level of diabetic retinopathy, we need to find and identify different types of lesions in the eye fundus. First, the normal anatomic structures of the eye (blood vessels, optic disc and fovea) must be removed from the image, in order to make visible the abnormalities. This thesis has focused on this step of image cleaning. This thesis proposes a novel framework for fast and fully automatic optic disc segmentation based on Markowitz's Modern Portfolio Theory to generate an innovative color fusion model capable of admitting any segmentation methodology in the medical imaging field. This approach acts as a powerful and real-time pre-processing stage that could be integrated into daily clinical practice to accelerate the diagnosis of DR due to its simplicity, performance, and speed. This thesis's second contribution is a method to simultaneously make a blood vessel segmentation and foveal avascular zone detection, considerably reducing the required image processing time. In addition, the first component of the xyY color space representing the chrominance values is the most supported according to the approach developed in this thesis for blood vessel segmentation and fovea detection. Finally, several samples are collected for a color interpolation procedure based on statistic color information and are used by the well-known Convexity Shape Prior segmentation algorithm. The thesis also proposes another blood vessel segmentation method that relies on an effective feature selection based on decision tree learning. This method is validated using three different classification techniques (i.e., Decision Tree, Artificial Neural Network, and Support Vector Machine).

Geographic object-based image analysis (GEOBIA) is an innovative image classification technique that treats spatial features in an image as objects, rather than as pixels; thus resembling closer to that of human perception of the geographic space. However, the process of a GEOBIA application allows for multiple interpretations. Particularly sensitive parts of the process include image segmentation and training data selection. The multiresolution segmentation algorithm (MSA) is commonly applied. The performance of segmentation depends primarily on the algorithms scale parameter, since scale controls the size of image objects produced. The fact that the scale parameter is unit less makes it a challenge to select a suitable one; thus, leaving the analyst to a method of trial and error. This can lead to a possible bias. Additionally, part from the segmentation, training area selection usually means that the data has to be manually collected. This is not only time consuming but also prone to subjectivity. In order to overcome these challenges, we tested a GEOBIA scheme that involved automatic methods of MSA scale parameterisation and training area selection which enabled us to more objectively classify images. Three study areas within Sweden were selected. The data used was high resolution Geografiska Sverigedata (GSD) orthophotos from the Swedish mapping agency, Lantmäteriet. We objectively found scale for each classification using a previously published technique embedded as a tool in eCognition software. Based on the orthophoto inputs, the tool calculated local variance and rate of change at different scales. These figures helped us to determine scale value for the MSA segmentation. Moreover, we developed in this study a novel method for automatic training area selection. The method is based on thresholded feature statistics layers computed from the orthophoto band derivatives. Thresholds were detected by Otsu’s single and multilevel algorithms. The layers were run through a filtering process which left only those fit for use in the classification process. We also tested the transferability of classification rule-sets for two of the study areas. This test helped us to investigate the degree to which automation can be realised. In this study we have made progress toward a more objective way of object-based image classification, realised by automating the scheme. Particularly noteworthy is the algorithm for automatic training area selection proposed, which compared to manual selection restricts human intervention to a minimum. Results of the classification show overall well delineated classes, in particular, the border between open area and forest contributed by the elevation data. On the other hand, there still persists some challenges regarding separating between deciduous and coniferous forest. Furthermore, although water was accurately classified in most instances, in one of the study areas, the water class showed contradictory results between its thematic and positional accuracy; hence stressing the importance of assessing the result based on more than the thematic accuracy. From the transferability test we noted the importance of considering the spatial/spectral characteristics of an area before transferring of rule-sets as these factors are a key to determine whether a transfer is possible.

L'anévrisme cérébral est une région fragile de la paroi d'un vaisseau sanguin dans le cerveau, qui peut se rompre et provoquer des saignements importants et des accidents vasculaires cérébraux. La segmentation de l'anévrisme cérébral est une étape primordiale pour l'aide au diagnostic, le traitement et la planification chirurgicale. Malheureusement, la segmentation manuelle prend encore une part importante dans l'angiographie clinique et elle est devenue couteuse en temps de traitement étant donné la gigantesque quantité de données générées par les systèmes d'imagerie médicale. Les méthodes de segmentation automatique d'image constituent un moyen essentiel pour faciliter et accélérer l'examen clinique et pour réduire l'interaction manuelle et la variabilité inter-opérateurs. L'objectif principal de ce travail de thèse est de développer des méthodes automatiques pour la segmentation et la mesure des anévrismes. Le présent travail de thèse est constitué de trois parties principales. La première partie concerne la segmentation des anévrismes géants qui contiennent à la fois la lumière et le thrombus. La méthode consiste d'abord à extraire la lumière et le thrombus en utilisant une procédure en deux étapes, puis à affiner la forme du thrombus à l'aide de la méthode des courbes de niveaux. Dans cette partie, la méthode proposée est également comparée à la segmentation manuelle, démontrant sa bonne précision. La deuxième partie concerne une approche LBM pour la segmentation des vaisseaux dans des images 2D+t et de l'anévrisme cérébral dans les images en 3D. La dernière partie étudie un modèle de segmentation 4D en considérant les images 3D+t comme un hypervolume 4D et en utilisant un réseau LBM D4Q81, dans lequel le temps est considéré de la même manière que les trois autres dimensions pour la définition des directions de mouvement des particules dans la LBM, considérant les données 3D+t comme un hypervolume 4D et en utilisant un réseau LBM D4Q81. Des expériences sont réalisées sur des images synthétiques d'hypercube 4D et d'hypersphere 4D. La valeur de Dice sur l'image de l'hypercube avec et sans bruit montre que la méthode proposée est prometteuse pour la segmentation 4D et le débruitage
Cerebral aneurysm is a fragile area on the wall of a blood vessel in the brain, which can rupture and cause major bleeding and cerebrovascular accident. The segmentation of cerebral aneurysm is a primordial step for diagnosis assistance, treatment and surgery planning. Unfortunately, manual segmentation is still an important part in clinical angiography but has become a burden given the huge amount of data generated by medical imaging systems. Automatic image segmentation techniques provides an essential way to easy and speed up clinical examinations, reduce the amount of manual interaction and lower inter operator variability. The main purpose of this PhD work is to develop automatic methods for cerebral aneurysm segmentation and measurement. The present work consists of three main parts. The first part deals with giant aneurysm segmentation containing lumen and thrombus. The methodology consists of first extracting the lumen and thrombus using a two-step procedure based on the LBM, and then refining the shape of the thrombus using level set technique. In this part the proposed method is also compared with manual segmentation, demonstrating its good segmentation accuracy. The second part concerns a LBM approach to vessel segmentation in 2D+t images and to cerebral aneurysm segmentation in 3D medical images through introducing a LBM D3Q27 model, which allows achieving a good segmentation and high robustness to noise. The last part investigates a true 4D segmentation model by considering the 3D+t data as a 4D hypervolume and using a D4Q81 lattice in LBM where time is considered in the same manner as for other three dimensions for the definition of particle moving directions in the LBM model

Thesis (M. S.)--Biomedical Engineering, Georgia Institute of Technology, 2009.
Committee Chair: Tannenbaum, Allen; Committee Member: Skrinjar, Oskar; Committee Member: Yezzi, Anthony. Part of the SMARTech Electronic Thesis and Dissertation Collection.

The focus of this thesis is on facial biometrics; specifically in the problems of face detection and face recognition. Despite intensive research over the last 20 years, the technology is not foolproof, which is why we do not see use of face recognition systems in critical sectors such as banking. In this thesis, we focus on three sub-problems in these two areas of research. Firstly, we propose methods to improve the speed-accuracy trade-off of the state-of-the-art face detector. Secondly, we consider a problem that is often ignored in the literature: to decrease the training time of the detectors. We propose two techniques to this end. Thirdly, we present a detailed large-scale study on self-updating face recognition systems in an attempt to answer if continuously changing facial appearance can be learnt automatically.
L'objectiu d'aquesta tesi és sobre biometria facial, específicament en els problemes de detecció de rostres i reconeixement facial. Malgrat la intensa recerca durant els últims 20 anys, la tecnologia no és infalible, de manera que no veiem l'ús dels sistemes de reconeixement de rostres en sectors crítics com la banca. En aquesta tesi, ens centrem en tres sub-problemes en aquestes dues àrees de recerca. En primer lloc, es proposa mètodes per millorar l'equilibri entre la precisió i la velocitat del detector de cares d'última generació. En segon lloc, considerem un problema que sovint s'ignora en la literatura: disminuir el temps de formació dels detectors. Es proposen dues tècniques per a aquest fi. En tercer lloc, es presenta un estudi detallat a gran escala sobre l'auto-actualització dels sistemes de reconeixement facial en un intent de respondre si el canvi constant de l'aparença facial es pot aprendre de forma automàtica.

碩士
國立中興大學
資訊科學與工程學系所
98
Malaria is an infectious parasitic disease which is widespread in Africa and South-East Asia. It is caused by a protozoan parasite of the genus Plasmodium. Since malaria parasites can be detected in the blood of infected patient, blood smear is commonly used to diagnose malaria under microscope. However, it is time-consuming and labor-intensive. Moreover, the experience of medical technologist has a great effect upon diagnostic accuracy. In this study, an automatic malaria parasite detector is proposed to perceive the malaria-infected erythrocytes from a blood smear image. This detector can more objectively and efficiently help the doctor diagnose malaria. Humans can be infected by four distinct species of Plasmodium parasites: Plasmodium falciparum (P. falciparum), Plasmodium vivax (P. vivax), Plasmodium ovale (P. ovale), and Plasmodium malariae (P. malariae). In this study, an automatic malaria parasite detector is proposed to diagnose the malaria from a blood smear image. The parasite appears in four stages in blood – ring, trophozoite, schizont, and gametocyte. Identifying the species and stage of the parasite is quite helpful in investigating the properties of malaria, preventing and diagnosing the malaria. Different Plasmodium species at different stage exhibit differences in their morphology and modify the host erythrocyte differently. It is hence feasible to develop an automatic system for identifying the species and the life stage of the parasite. How to segment the infected erythrocytes from a blood smear image and the parasite from the infected erythrocytes is essential for developing the automatic system. The other tasks of this study are hence to cut off the infected erythrocytes and parasite. The experimental results tell that the proposed method can provide impressive performance in detecting the malaria-infected erythrocytes and segmenting the infected erythrocytes and parasites.

碩士
國立雲林科技大學
資訊管理系碩士班
99
Recently, high prevalence of lung cancer, more and more researcher concern about diagnosing pulmonary lesions in chest CT images. Chest computed tomography (CT) is a well-established tool, which is widely applied to related detection work. However, the great amount of CT scans of each patient still is a challenge for specialists. Hence, many literatures have proposed methods for automatic diagnosis by computer-aided detection (CAD) to assist artificial inspection. But there are some problems brought with those proposed methods, such as time-consuming work of automatically diagnosing by CAD for the large number of CT slices. Even though a patient has trouble with pulmonary nodule positively, that not means there are nodules in each chest slice of him. Therefore this study proposed a hybrid method to initially classify lungs images into 3 classes: nodule, non-nodule, inflammation. The proposed method can be executed before doctor diagnosis or computer-aided system, which can be sure that input CT image need to be detected out the actual positions, shapes or other information of nodules. The results display a higher accuracy in RS based on WP-SVD than other classification methods, which verifies that proposed method can reduce time and cost of lung nodule diagnosis.

碩士
國立中興大學
資訊管理學系所
102
This paper is based on SMRIVS method to refine the results of SMRIVS method by eliminating thecal sac, intervertebral disc and other tissues on a MRI image. The proposed method is divided into three stages. The first stage is preprocessing stage, which adjusts the intensities of the results of SMRIVS method for further use. The purpose of the second stage is to eliminate intervertebral disc and some adjacency darker tissues. Because the intensities of those parts are darker than those of the vertebra bone, local thresholding and monograph operators to eliminate those parts. After that, the third stage is used to thecal sac. Thecal sac parts have high intensity in MRI image because it’s full of cerebrospinal fluid. Therefore, the intensities of the results of the second stages are enhanced using gamma equalization. The thecal sac can be eliminated by perform otsu’s method on the enhanced images. Finally, the spinal set from the MRI image can be obtained successfully. The average segmentation performance of the segmentation results in 36 spinal MRI images is 83.21%.

碩士
國立中山大學
資訊工程學系研究所
92
Recently, for the new demands of nowadays multimedia system, such as video interaction, the MPEG-4 standard has been designed. In MPEG-4, because of those new demands of nowadays multimedia system the video stream can be divided into several video object planes ( VOPs ). Those VOPs can be separately encoded, stored, or transmitted. VOP is the basic interactive unit in MPEG-4 video stream, how to automatically or semi-automatically separate appropriate VOPs from an image sequence has become one of the most important issues for an MPEG-4 system, which is also the goal of this proposal. However, MPEG-4 does not provide concrete techniques for VOP extraction. Nonetheless, it is very difficult to extract VOPs, thus the preprocessing used to decompose sequences into VOPs becomes an important issue for an MPEG-4 system, which is also the goal of this thesis. In this thesis, we will develop techniques for segmenting images contained in an image sequence, which can separate two or more image segments ( or regions ) from MPEG-4 test image sequences, and those image segments can be coded as MPEG-4 VOPs. First, we utilize the feature of wavelet to improve the change detection, such that we can obtain a better result of the moving object edge by improved change detection. Second, we use an edge-based method for tracking boundary which is using the canny edge detection and the connected edge component labeling to label those edges. Third, we can combine those two information to obtain a more complete boundary by extracting moving object edges. Although we catch all the edges which is detected on the location of the true boundary, it usually occurs some gaps on which we catch. Because it sometimes will not have a clear boundary, we have to find some method to complete these gaps. Therefore, we propose a multi-level prediction scheme to complete the gaps between the disjoint edges of the boundary we caught by extending the edges on the predictive direction. Final, we use a simple connecting operation for the little gaps (distance=1 or 2). That will make the result more close and smooth. Experimental results for several test sequences show that this novel automatic video segmentation algorithm can give a more accurate object masks.

碩士
中台醫護技術學院
醫學工程暨材料研究所
95
Recently, active contour model (ACM) algorithm has been widely used to perform image segmentation in many clinical imaging modalities. It has been proven to be an efficient and accurate imaging segmentation method to segment the region of interest in high noise or low contrast image. In addition to segment out the lesion contour, how to accurately quantify the volume of the lesion in an image is very important. The direct counting method has been used to quantify the volume of ACM segmented regions. Although the error of the direct counting method is quite small in some segmented regions with simple shapes, the error is quite large in most segmented region with some concave or convex shapes. In order to alleviate the problems encountered by direct counting method, a new volume quantification method were proposed. The purpose of this study is to develop an accurate and efficient automatic volume quantification method for the ACM segmented region. The method presented herein includes edge linking and region filling processes. Various types of computer simulated images were created to evaluate the accuracy of the automatic volume quantification method. In the mean while, the relationship between image matrix size and the quantification of ACM segmented region were also investigated. These computer simulated images were categorized into three types, closed contour, discrete contour and regional types. Images with closed contour were used to evaluate the accuracy of the region filling process. The results showed that the accuracy of region filling process was 100%, and the average percentage error of direct counting method was 15.85%. It means that region filling process is more accurate and more reliable than direct counting method in quantifying the segmented region. Images with discrete contour were used to evaluate the accuracy of the edge linking process. The results showed that the average percentage error of edge linking and region filling process was 0.03% which is much smaller than that of direct counting method 15.08%. It means that edge linking and region filling process is more accurate and more reliable than direct counting method in quantifying the segmented region. Images with regional type were used to evaluate the accuracy of the ACM image segmentation. The results showed that the average percentage error of edge linking and region filling process after ACM image segmentation process was 7.52% which was much smaller than that of direct counting method 20.36%. It means that edge linking and region filling process after ACM image segmentation process is more accurate and more reliable than direct counting method in quantifying the segmented region. Regional type images with various matrices were used to evaluate the correlation between image matrix size and the quantification of ACM segmented region. The results showed that the average percentage errors of edge linking and region filling process after ACM image segmentation process were 17.55%(256 × 256), 10.89%(512 × 512) and 6.21%(1024 × 1024). It means that the larger the image matrix, the smaller the average percentage error. The larger the image matrix, the higher the accuracy of ACM image segmentation process. From the point of view of morphology, it also showed that the larger the image matrix of clinical MR images, the higher the accuracy of ACM image segmentation process and volume quantification. The proposed volume quantification method can be used in quantifying the volume of segmented region of interest after ACM image segmentation process accurately. It is capable of quantifying the volume of malignant lesion with irregular shapes. It might be useful to increase the efficiency, accuracy, externality in clinical diagnosis and the evaluations of pre-surgical, post-surgical or treatment plan.

碩士
國立中正大學
資訊工程研究所
107
In the procedures of radiotherapy, delineating the organ at risk (OAR) is a time-consuming, laborious but still very important task, which is necessary to be done accurately. In the field of medical imaging analysis, the content of images has high complexity and variability. The traditional rule-based methods are often difficult to meet the clinical requirements. In this work, we study the deep learning segmentation techniques and apply them to clinical imaging systems. In the chapter of experimentation, we used public liver organ datasets, 3Dircadb, to verify various types of segmentation models. By the inspiration of these models we proposed a fusion model that uses the Convolutional LSTM Layer to study the spatial correlation between layers in a CT image dataset. We also use Attention Mechanism to suppress irrelevant features from the complex image content and focus on the useful messages of target organs. Finally, this model is verified in the testing dataset and achieves the highest Dice score. In addition, the 2015 MICCAI public data set on organ segmentation was used to segment the stomach. We show that the proposed fusion model still has the best performance even when the organ boundaries are more difficult to discriminate. The last experimental results show that the diversity of training dataset used by deep learning techniques is very important in clinical application. For patients with special disease conditions, extra data with similar characteristics is required to make the prediction more accurate. Pre-trained model parameters can improve the results of testing, but the results are poor when directly applied to test a different dataset. We also proposed three feasible solutions for practically applying deep learning methods in radiotherapy. It allows us to successfully write the predicted contour results into DICOM-RT format, which is a standard data format compatible with most medical imaging systems. Therefore the clinician could directly fine-tune the predicted contour and save the time for other clinical work.

碩士
國立臺灣大學
生醫電子與資訊學研究所
100
In recent years, the computer-aided diagnostic (CAD) is developed gradually providing second opinion for radiologists'' diagnosis. The CAD system should produce well segmentation result of tumor to precisely evaluate the tumor size, and the tumor can be further classified into benign and malignant by the extracted feature of shape. Concerning the previous techniques of the semi-automatic segmentation (e.g., level-set), the seed usually needs to be manually initialized at the appropriate position for the better segmentation result. Besides, most of the segmentation systems consider the detection procedure as the preceding step, and the segmentation approach is subsequently employed on the located region of interest. In this study, a fully automatic system integrating the tumor detection and segmentation steps is proposed, and a set of representative seeds are computerized for the whole image segmentation based on the level-set method. Meanwhile, the novel multi-seed mechanism assists the level set segmentation in acquiring the satisfactory result. The proposed system consists of three phases. First, a mean-shift clustering method and the affinity approach are applied to generate and extract the most representative seeds. Next, the level-set method based on the selected seeds is employed to obtain several suspected regions. Finally, the features of all the suspected regions are calculated and further analyzed by support vector machine (SVM) classifier to extract the target tumor from the normal ones. 120 cases (68 for benign cases and 52 for malignant cases) are used for evaluating the proposed system. The sensitivity is 95% (114/120) with the benign cases equal to 0.91 and the malignant cases to 1.00. In summary, the experimental results demonstrate the high efficiency and robustness of the proposed method.

碩士
國立臺中科技大學
資訊管理系碩士班
102
This study proposes a new method for medical image segmentation, automatic Level set cutting method, the liver from computed tomography (CT) images quickly and accurately cut out. Automatic liver cut method of this study will be divided into a preliminary outline of the liver and liver Level set to detect cut. Liver preliminary outline detection system using a dedicated medical images in DICOM format Hounsfield Unit (HU), extracting the liver window width and window level to set the threshold, and morphological techniques to complete the initial liver contour cutting. Cut liver Level set method takes advantage liver contour Level set of initial treatment to obtain more precise liver profile. Experiments show that the proposed method to detect liver accuracy rate is very high, and cutting accuracy rate as high as 95.23% of the liver. This method can improve the accuracy and reduce the liver cut processing time, full use of DICOM format images to be cut into convenient in-hospital PACS system, hoping to assist interpretation of the practical applications for use in clinical physician.

碩士
國立中興大學
應用數學系所
94
In the paper proposes a New Adaptive Threshold Level Set Model. The model is based on the methodology of Active Contours Without Edges . The proposed method is then applied to the segmentation of gray matter in brain MR images. There are two threshold values, lower threshold and upper threshold, to be decided for locating the contours of the gray matter of the brain MR images. The threshold values are determined by Fuzzy C-Means Algorithm. By combining segmentation of 134 planar MR brain gray images. A three-dimensional model of the brain gray matter has been successfully constructed.

Programa de Doutoramento em Informática (MAP-i) das Universidades do Minho, de Aveiro e do Porto
Gliomas are the most common primary brain tumors. Unfortunately, these neoplasms hold the worst prognosis among all brain tumors, as well. They can be broadly categorized as low or high grade gliomas. Magnetic Resonance Imaging is the standard imaging technique for their assessment. Using it, physicians can extract measurements that are crucial for treatment planning and follow-up. Notwithstanding, manual segmentation is time-demanding and prone to variability. Also, tumor grading by biopsy is very important, but it is invasive, and prone to sampling error. Therefore, automatic approaches for both segmentation and grading are needed. However, these tasks are quite challenging due the the heterogeneity of gliomas, as well as the variability among Magnetic Resonance Imaging scans. This makes it difficult to model brain tumors from prior knowledge. Machine Learning algorithms can learn how to perform a task directly from the data. Some of these algorithms may be categorized as Representation Learning if they can learn features directly from the data. Among these methods, Deep Learning is a group of Representation Learning algorithms that learn multiple levels of representations. In the past years, Deep Learning-based methods have shown remarkable performances. Hence, the aim of this work was to investigate Deep Learning methods, and use them for the automatic segmentation and grade classification of brain tumors in multisequence structural Magnetic Resonance Imaging. Additionally, an often cited setback of these complex models is their “black box” behavior. Thus, in this work we also studied interpretability of Machine Learning algorithms applied to medical imaging. Therefore, we built our work on: brain tumor image analysis in Magnetic Resonance Imaging, Machine Learning with focus on Representation Learning, and its interpretability. We investigated Convolutional Neural Networks for the task of segmentation. As a starting point, we studied a classification Convolutional Neural Network. We were able to show its effectiveness, as well as the importance of careful pre-processing. However, afterwards we adopted a more efficient Fully Convolutional Network approach. In this setting, we proposed a hierarchical approach for dealing with class imbalance. Finally, the relationships among channels of feature maps were studied. We proposed and showed the benefits of recombination and recalibration of feature maps in the context of Fully Convolutional Networks for semantic segmentation. Automatic glioma grading from structural Magnetic Resonance Imaging images is challenging due to their large heterogeneity. Additionally, a tumor mass must be graded as a whole. Therefore, we propose 3D Convolutional Neural Networks for automatic glioma grading. Since Convolutional Neural Networks learn features directly from the data, it allows one to bypass the need for a very accurate segmentation that is often seen in radiomics-based approaches, which use hand-crafted features. “Black box” systems may pose trusting issues when deployed in critical domains, such as the medical field. This is due to professionals not being able to explain certain predictions. Therefore, interpretability of machine learning systems is a crucial field of research, given the high performances currently achieved with these systems. We first investigated this topic in a Restricted Boltzmann Machine and Random Forest classifier system in the context of segmentation. We proposed methodologies for both global and local interpretability. The former is targeted at understanding if the system learned the relevant relations in the data, while the latter is focused on explaining individual predictions. We were able to confirm if the system learned correct patterns, but we also found a bias in the database. Later, we employ interpretability methodologies to inspect the 3D Convolutional Neural Network for glioma grading. With it, we were able to catch and correct an issue during pre-processing. Hence, we provide tools and study cases that show how interpretability not only helps in increasing trust, but it may also be useful during the development cycle. Finally, all methodologies developed in this work were validated in publicly available databases. This ensures a fair comparison with the state of the art. Additionally, it enables future work to be directly compared with us.
Os gliomas são os tumores cerebrais primários mais frequentes. Infelizmente, estas neoplasias são também as que que têm os priores prognósticos entre os tumores cerebrais. Estes podem ser categorizados em gliomas de baixo ou de alto grau. A Imagem por Ressonância Magnética é a técnica imagiológica padrão para avaliar tumores cerebrais. Desta forma, os médicos podem extrair medições que são da maior importância para o planeamento do tratamento e para monitorização. Não obstante, a segmentação manual das imagens é um processo demorado e suscetível a variabilidade. A classificação dos gliomas quanto ao seu grau através de biópsia é também muito importante, mas é um processo invasivo, e suscetível a erros de amostragem. Assim sendo, são necessárias abordagens automáticas para ambas as tarefas. Contudo, são problemas bastante complexos devido à heterogeneidade dos gliomas, mas também devido à variabilidade das imagens de Ressonância Magnética. Isto faz com que seja difícil modelar os tumores cerebrais a partir de conhecimento a priori. Os algoritmos de Aprendizagem Automática conseguem aprender a executar uma determinada tarefa diretamente a partir dos dados. Alguns destes algoritmos podem ser categorizados como Aprendizagem de Características se forem capazes de aprender características diretamente a partir dos dados. Entre estes métodos, Aprendizagem Profunda é um grupo de algoritmos de Aprendizagem de Características que aprendem múltiplos níveis de características. Nos últimos anos, métodos baseados em Aprendizagem Profunda têm mostrado desempenhos notáveis. Assim, um dos objetivos deste trabalho foi a investigação de Aprendizagem Profunda no contexto de segmentação. Um segundo objetivo foi explorar Aprendizagem Profunda para a classificação automática dos graus dos gliomas a partir de Imagem por Ressonância Magnética estrutural. Finalmente, um problema que é muitas vezes apontado a estes modelos complexos é a sua natureza de “caixa preta”. Assim sendo, neste trabalho também foi investigada a interpretabilidade de sistemas de Aprendizagem Automática. Portanto, há três grandes temas sobre os quais nós construímos o nosso trabalho: análise de imagem de tumores cerebrais em Imagem por Ressonância Magnética, Aprendizagem Automática com foco em Aprendizagem de Características, e interpretabilidade. Foram investigadas Redes Neuronais Convolucionais para a tarefa de segmentação. Como ponto de partida, nós estudamos Redes Neuronais Convolucionais para classificação. Assim, conseguimos mostrar a sua eficácia, tal como a importância de um pré-processamento cuidadoso. Contudo, posteriormente, nós adotamos uma abordagem mais eficiente denominada por Redes Totalmente Convolucionais. Com esta nova rede, nós propusemos uma abordagem hierárquica que nos permitiu lidar melhor com o desbalanceamento de classes. Por fim, as relações entre os canais dos mapas de características foram estudadas. Nós propusemos e mostrámos as vantagens da recombinação e recalibração dos mapas de características no contexto de Redes Totalmente Convolucionais para segmentação semântica. A classificação automática do grau dos gliomas a partir de Imagem por Ressonância Magnética estrutural é complexa devido à sua grande heterogeneidade. Adicionalmente, uma massa tumoral deve ser classificada como um todo. Assim sendo, nós propusemos Redes Neuronais Convolucionais 3D para a classificação automática do grau dos gliomas. Uma vez que as Redes Neuronais Convolucionais aprendem características diretamente a partir dos dados, é possível evitar a necessidade de uma segmentação muito precisa, tal como é comumente observado nas abordagens mais tradicionais baseadas em radiomics. Os sistemas “caixa preta” podem colocar problemas relacionados com confiança quando são colocados em domínios críticos, como é o caso do campo da medicina. Isto deve-se aos profissionais não serem capazes de explicar certas predições. Assim sendo, a interpretabilidade de sistemas de Aprendizagem Automática é uma área de investigação crucial, dado os elevados desempenhos atualmente atingidos com estes sistemas. Nós investigamos este tópico inicialmente com um sistema constituído por uma Restricted Boltzmman Machine e um classificador Random Forest no contexto de segmentação. Nós propusemos metodologias para interpretabilidade global e local. A primeira está direcionada para se perceber se o sistema aprendeu relações relevantes nos dados, enquanto a última foca-se mais em explicar predições individuais. Nós conseguimos confirmar se o modelo aprendia padrões corretos, mas também conseguimos encontrar um enviesamento na base de dados. Posteriormente, também aplicamos metodologias de interpretabilidade para inspecionar a Rede Neuronal Convolucional 3D para classificação do grau dos gliomas. Assim, conseguimos identificar e corrigir um problema durante o pré-processamento. Desta forma, nós fornecemos ferramentas e casos de estudo que mostram como a interpretabilidade é útil não só para aumentar a confiança, mas também durante o ciclo de desenvolvimento. Finalmente, todas as metodologias desenvolvidas neste trabalho foram validadas em bases de dados publicamente disponíveis. Isto garante uma comparação justa com o estado da arte. Adicionalmente, isto permite que trabalhos futuros possam ser diretamente comparados com os nossos métodos.
This work was supported by a scholarship awarded to Sérgio Rafael Mano Pereira by Fundação para a Ciência e Tecnologia (FCT), Portugal, with scholarship reference PD/BD/105803/2014.

碩士
高雄醫學大學
醫學影像暨放射科學系碩士在職專班
105
Background and Objective: Magnetic Resonance Imaging (MRI) has been previously utilized to evaluate the extent of muscle mass loss by manual drawing of muscle area which however is prone to subjective errors. The purpose of this study was to investigate the correlation between the bone mineral density (BMD) obtained from dual-energy X-ray absorptiometry (DXA) and the muscle area obtained from MRI using semi-automatic segmentation. Materials and Methods: This study reviewed 85 postmenopausal women who were older than 50 years old and had undergone both DXA and MRI examinations in 3 months. Of 85, 50 patients were without osteoporosis and 35 patients were with osteoporosis. None of them had compression fracture, osteophyte, severe scoliosis, bone surgery, metal implant, and dialysis. This study performed a semi-automatic Gaussian mixture segmentation method to obtain lean muscle area of erector spinae from five axial spine MRIs (L1-L2, L2-L3, L3-L4, L4-L5, L5-S1), and used Pearson’s correlational analysis to investigate the correlations between BMD, lean muscle area, and other parameters. Results: Correlational analysis showed that in 85 patients, the BMD was significantly correlated with muscle area, body weight, BMI, age, and duration of amenorrhea, and the L1-L2 muscle area was found to have higher correlation with BMD than those of lower slices. In non-osteoporotic group, the muscle area was significantly correlated with BMD only at L1-L2 slice. However, in osteoporotic group, the muscle areas of most slices were significantly correlated with BMD, and the muscle area at L1-L2 level had higher correlation than other levels. Discussion and Conclusion: This study demonstrated that the semi-automatic segmentation method is helpful to accurately measure the core muscle area which can be used to assess the severity of osteoporosis in postmenopausal women. The results also showed that the muscle area at L1-L2 level had higher significant correlation with BMD than other slices, suggesting that upper slice is more perpendicular to the erector spinae than lower slices and the obtained muscle cross-sectional area is more accurate. Therefore, we concluded that L1-L2 muscle area is suitable for investigating the severity of osteoporosis in postmenopausal women, not only because it has higher correlation with BMD but also because it takes less image analyzing time. Keywords: BMD, DXA, MRI, Sarcopenia, Postmenopausal Women, Osteoporosis, Core muscles

Road networks play an important role in a number of geospatial applications, such as cartographic, infrastructure planning and traffic routing software. Automatic and semi-automatic road network extraction techniques have significantly increased the extraction rate of road networks. Automated processes still yield some erroneous and incomplete results and costly human intervention is still required to evaluate results and correct errors. With the aim of improving the accuracy of road extraction systems, three objectives are defined in this thesis: Firstly, the study seeks to develop a flexible semi-automated road extraction system, capable of extracting roads from QuickBird satellite imagery. The second objective is to integrate a variety of algorithms within the road network extraction system. The benefits of using each of these algorithms within the proposed road extraction system, is illustrated. Finally, a fully automated system is proposed by incorporating a number of the algorithms investigated throughout the thesis. Copyright
Dissertation (MSc)--University of Pretoria, 2010.
Computer Science
unrestricted

碩士
國立臺北科技大學
自動化科技研究所
97
This article presents an innovative image segmentation method to extract an object underlying defect detection from its background image. 2-D automatic optical inspection (AOI) technology for defect detection and classification has played a vital role for in-situ manufacturing industrial sectors nowadays. Image segmentation is a crucial step to extract component information from its neighboring background. Due to potential complexity in such an image processing operation, considerable challenges are crucially encountered in establishing a robust approach. In general, three major factors play a significant influence on the result of the segmented image objects: (1) brightness distribution of the background image; (2) degree of unbalanced brightness of the background image; (3) noise level near the object feature to be detected. The research addresses these important factors and develops an effective segmentation method. Exclusive advantage of the method is to overcome the current limitations of the existing SVD (singular value decomposition) or DCT (discrete Cosine transfer) methods. The segmentation performance of the developed method is up to 14% better than the DCT method, in terms of accuracy of segmentation. From the test results on some real industrial cases, it is verified that the method is capable of extracting the tested object desirably.

碩士
國立清華大學
電機工程學系
104
Lesion segmentation for breast ultrasound images has been studied by many people, but it is hard to achieve using traditional edge detection because ultrasound images don’t have sharp contours and exist a lot of noise. In this thesis, we propose an automatic segmentation method for breast ultrasound images which combines morphological image processing and distance-regularized level-set evolution method (DRLSE), and improve combined DRSLE (cDRLSE) proposed by Yung-Hsuan Hsu. The most serious problem of cDRLSE is that it spend a lot of time on obtaining initial contour through the operation which applied the texture features (e.g., gray level co-occurrence matrix (GLCM)) for support vector machine (SVM), therefore we propose a new method based on morphological operations to search for initial contour which is effective and efficient. Using the new initial contours, we can capture the tumor area more precisely after applying DRLSE. To evaluate the result of segmentation, we compare it with expansion DRLSE method, contraction DRLSE method and cDRLSE method using three evaluation metrics, including misclassification error (ME), relative foreground area error (RFAE) and modified Hausdorff distance (MHD). We find that the proposed method is basically better than expansion DRLSE method and contraction DRLSE method which confirms the importance of initial contour to DRLSE. However, it is better than cDRLSE method in RFAE but worse than cDRLSE in ME and MHD, probably due to ranking failure in the proposed method or improved segmentation accuracy of post processing in cDRLSE method; even so, the proposed method not only spends less time obviously but also has no need to apply post processing. The proposed method has the following properties: 1.A fully automatic segmentation method for breast ultrasound images which has no need to set initial contour manually. 2.The way to obtain initial contour is efficient, and moreover if the tumor is smooth, the initial contour will be close to the tumor’s real boundaries. 3.Compared with other methods, the segmentation result of the proposed method is truly closer to ground truth image if the proposed initial contour mostly lies inside the ground truth.

碩士
中原大學
電機工程研究所
98
Segmentation of magnetic resonance (MR) images has numerous clinical applications: Being an auxiliary tool in diagnosis and treatment, making epidemiological statistics easier to carry on, being an important step in data analysis of all kinds of medical researches. However, most of the segmentation has been done manually or at most semi-automatically. The process is time- and energy-consuming and difficult to be standardized. Automatic segmentation algorithms of MR images for some of the body parts, such as the brain, have been developed with success. The muscular tissue is a major component of the human body, however, to our knowledge, no similar studies had been done on automatic muscle segmentation on MR images. The goal of this study is to develop an automatic segmentation scheme to correctly assign different functional muscle groups on MR images of the human lower extremities. We speculated that the results could be used in increasing muscle-tissue-related researches, such as the monitoring of muscle volume change over the time for the victims of muscular dystrophy diseases and investigation of the use and the performance of different muscles in athletes of various sport types. The grouping and the surrounding anatomic structures are quite different for muscles located at different body parts. For convenience and to support another research of our labs, we chose to target at lower extremities, muscles from pelvis to ankle, as the object of the study. Since the MR signals for all muscles are more or less similar, instead of singling out an individual muscle, our study focused on automatic classification of functional muscle groups. In the thesis, the lower extremities were first divided into several longitudinal anatomic segments based on the principles of proximity and anatomic similarity. Since the MR signals are quite different among bones, fascia, fat, and muscles, subcutaneous fat, ilium of the pelvis, femur, tibia, fibula, meniscus, and the muscular fascia are available as boundaries. Equipped by this knowledge, we applied mathematical morphological operations such as erosion, dilation, open, and close, and other image processing techniques such as regional growing and filling to designate muscle areas on an MR image as one of the eight muscle functional groups. The results of the automatic segmentation were then compared against the classification manually made by a physical therapist. We found that the average total absolute error (relative error) ± the standard deviation of our automatic segmentation is 1736.5 (13.5%) ± 1071.8 ml, with 490.2 (14.0%) ± 371.8 ml of the knee extensors as the largest absolute error and 41.3 (7.7%) ± 11.6 ml of the ankle flexors as the smallest. On the other hand, the largest relative error is hip flexors’ 242.8 (24.4%) ± 140.2 ml and the smallest the ankle extensors’ 153.1 (6.7%) ± 94.3 ml. The study presents that the combined mathematical morphology and human anatomy knowledge approach successfully divided muscles of lower extremity MR images into meaningful functional groups without human intervention. In the future, the accuracy of this method could be further improved by more sophisticated revision such as MR-atlas registration. Applications on other body parts and tissues such as abdominal visceral fat are under investigation. We expect the results of this and related studies to be helpful in body-composition-related researches and perhaps also in clinical diagnosis and treatment.

碩士
國立清華大學
電機工程學系所
105
Due to the speckle noise and low contrast in breast ultrasound images, it is hard to locate the contour of the tumor by using a single method. In this thesis, a new method for finding an initial contour is proposed, which can improve the result of DRLSE on the segmentation of BUS images. The new method focuses on improving the algorithm proposed by Tsai-Wen Niu, which is a way to search an initial contour based on the local minimum in the images. When the BUS images contain calcification, it is possible to fail in searching of initial contour through such algorithm, hence leading to a poor segmentation result when the initial contour is on the wrong place. Therefore, we acquire a bigger initial contour by using a series of image smoothing methods and binarization, which can eliminate the weak edges and adjust the contrast in BUS images. In addition, some images without local minimum can be successfully detected by using the proposed method. However, the pixel value in these images are similar. It might be hard to accurately separate the tumor region from non-tumor region by the difference of pixel values. These obstacles are conquered by calculating the difference of length and pixel value in the suspect region. The ranking outcome is improved by using the morphological features. After applying DRLSE, our initial contour can reach the tumor region more accurately. To evaluate the result of segmentation, it is compared with the outcome of DRLSE obtained from different initial contours proposed by Tsai-Wen Niu, expansion DRLSE method, and contraction DRLSE method using three evaluation metrics, including ME, RFAE and MHD. The experimental results indicate that the proposed method is basically better than the other methods. However, the initial contour might contain non-tumor region when the edge of the tumor’s boundary is too ambiguous; even so, the proposed method drastically reduce the number of DRLSE iteration and computation time. According to the experimental results, the proposed method has three advantages over the other methods. First, it sets the initial contour automatically which is more efficient than setting the initial contour manually. Second, the region of the initial contour is much bigger than those obtained by the other methods, which can reduce the computation time and the number of DRLSE iteration. Third, if the tumor boundary is distinct, the new initial contour can improve the segmentation result of DRLSE.