Dissertations / Theses on the topic 'Computer Aided Diagnosis'

Automatic analysis and information extraction from an image is still a highly chal- lenging research problem in the computer vision area, attempting to describe the image content with computational and mathematical techniques. Moreover the in- formation extracted from the image should be meaningful and as most discrimi- natory as possible, since it will be used to categorize its content according to the analysed problem. In the Medical Imaging domain this issue is even more felt because many important decisions that affect the patient care, depend on the use- fulness of the information extracted from the image. Manage medical image is even more complicated not only due to the importance of the problem, but also because it needs a fair amount of prior medical knowledge to be able to represent with data the visual information to which pathologist refer. Today medical decisions that impact patient care rely on the results of laboratory tests to a greater extent than ever before, due to the marked expansion in the number and complexity of offered tests. These developments promise to improve the care of patients, but the more increase the number and complexity of the tests, the more increases the possibility to misapply and misinterpret the test themselves, leading to inappropriate diagnosis and therapies. Moreover, with the increased number of tests also the amount of data to be analysed increases, forcing pathologists to devote much time to the analysis of the tests themselves rather than to patient care and the prescription of the right therapy, especially considering that most of the tests performed are just check up tests and most of the analysed samples come from healthy patients. Then, a quantitative evaluation of medical images is really essential to overcome uncertainty and subjectivity, but also to greatly reduce the amount of data and the timing for the analysis. In the last few years, many computer assisted diagno- sis systems have been developed, attempting to mimic pathologists by extracting features from the images. Image analysis involves complex algorithms to identify and characterize cells or tissues using image pattern recognition technology. This thesis addresses the main problems associated to the digital microscopy analysis in histology and haematology diagnosis, with the development of algorithms for the extraction of useful information from different digital images, but able to distinguish different biological structures in the images themselves. The proposed methods not only aim to improve the degree of accuracy of the analysis, and reducing time, if used as the only means of diagnoses, but also they can be used as intermediate tools for skimming the number of samples to be analysed directly from the pathologist, or as double check systems to verify the correct results of the automated facilities used today.

Les outils d’aide au diagnostic sont de nos jours au cœur de plusieurs protocoles cliniques car ils améliorent la qualité du diagnostic posé et des soins médicaux. Ce travail de recherche met en avant une architecture hiérarchique pour la conception d'un outil d'aide à la détection du cancer du sein robuste et performant. Il s’intéresse à la réduction des fausses alarmes en identifiant les régions potentiellement cancérogènes. La gamme dynamique des niveaux de gris des zones sombres est étirée pour améliorer le contraste entre la région du sein et l'arrière plan et permettre une meilleure extraction de celle-ci. Toutefois, le muscle pectoral demeure incrusté dans la région du sein et interfère avec l'analyse des tissus. Son extraction est à la fois difficile et complexe à mettre en œuvre à cause de son chevauchement avec les tissus denses du sein. Dans ces conditions, même en exploitant l'information spatiale pendant la clusterisation par un algorithme de fuzzy C-means ne produit pas toujours des résultats de segmentation pertinents. Pour s'affranchir de cette difficulté, une étape de validation suivie d'un ajustement de contour est mise sur pied pour détecter et corriger les imperfections de segmentation. La seconde étape est consacrée à la caractérisation de la densité des tissus. Pour faire face au problème de variabilité des distributions de niveaux de gris dans les classes de densités, nous introduisons une modification de contraste basée sur un transport optimisé de niveaux de gris. Grâce à cette technique, la surface relative de tissus denses estimée par simple segmentation est très fortement corrélée aux classes de densités issues d’un jeu de données étiquetées
Computer-aided diagnosis systems are currently at the heart of many clinical protocols since they significantly improve diagnosis making and therefore medical care. This research work therefore puts forward a hierarchical architecture for the design of a robust and efficient CAD tool for breast cancer detection. More precisely, it focuses on the reduction of false alarms rate through the identification of image regions of foremost interest i.e potential cancerous areas. The dynamic range of gray level intensities in dark regions is, first of all stretched to enhance the contrast between tissues and background and thus favors accurate breast region extraction. A second segmentation follows since pectoral muscle which regularly tampers breast tissue analysis remains inlaid in the foreground region. Extracting pectoral muscle tissues is both hard and challenging due to its overlap with dense tissues. In such conditions, even exploiting spatial information during the clustering process of the fuzzy C-means algorithm does not always produce a relevant segmentation. To overcome this difficulty, a new validation process followed by a refinement strategy is proposed to detect and correct the segmentation imperfections. The second macro-step is devoted to breast tissue density analysis. To address the variability in gray levels distributions with of mammographic density classes, we introduce an optimized gray level transport map for mammographic image contrast standardization. Thanks to this technique, dense region areas computed using simple thresholding are highly correlated to density classes from an annotated dataset

The research activity conducted during my PhD aims to develop two different Computer Aided Diagnosis (CAD) systems for breast and prostate cancer diagnosis using Magnetic Resonance Imaging. During the first part of this thesis I will illustrate a fully automatic CAD system for breast cancer detection and diagnosis with Dynamic Contrast Enhanced MRI (DCE-MRI) developed by our group. The main goal of a CAD system is lesions detection and characterization. The processing pipeline includes automatic segmentation of the breast and axillary regions, registration of unenhanced and contrast-enhanced frames, lesion detection and classification according to kinetic and morphological criteria. During my PhD I, firstly, studied the physiological phenomena correlated to breast tumors growth and diagnosis, then I elaborated and created C++ algorithms for: 1. breasts segmentation, where the breasts and axillary regions are automatically identified in order to reduce the computational burden and preventing false positives (FP) due to enhancing structures (such as the heart and extra-breast vessels) which are not of clinical interest. 2. lesion detection, in which suspicious areas showing contrast enhancement are automatically segmented and FPs are identified and discarded. These step are innovative as they are fully automatic, thus they do not suffer of inter- and intra-operator variability, and because of the normalization process, based on the mammary arteries segmentation, that makes the system able to deal with images coming from different centers, thus having different acquisition parameters. The second part of my thesis will concern the development of a CAD system for prostate cancer. The importance of this project is associated to the recent interest in adapting focal methods of tissue ablation, such as cryotherapy and Focused Ultrasound guided by MR (MRgFUS), to cure or control localized prostate cancer. Focal treatments rely on imaging to locate tumor, to determine the staging of disease, to detect recurrences and to guide the treatment. The aim of this part of my PhD was to create a multispectral computer aided diagnosis system able to: a) detect the tumor in order to guide real-time biopsy, b) characterize the malignancy of the lesion and c) guide the local treatment, by adopting a new multispectral approach. In this project I, actively, elaborated and developed C++ algorithm to register different datasets and to monitor the focal treatment using Diffusion Weighted-MRI (DWI) self-made acquisitions.The registration between T2-w, DCE-MRI and DWI images are applied in order to correct for patients movements and DWI distortions. Results obtained within 19 patients showed a Dice’s overlap coefficient higher than 0.7, considered optimal in literature. Monitoring the focal therapy was the aim of the last part of this project, that I actively developed during a visiting period in the Radiological Science Laboratory of the Stanford University (Kim Butts Pauly Research Lab). The main goal was to characterize the role of the DWI during MRgFUS. DWI, in fact, is very sensitive to cell death and tissue damage and information can be used to evaluate the treatment without relocating the patient and the applicators and without involving the administration of contrast agent. In this study, I wanted to assess the use of DWI images to estimate prostate tissue damage during HIFU ablation, by measuring diffusion coefficients of canine prostate pre and post ablation, using multiple b-factors ranging up to 3500 s/mm2 . This study demonstrated a bi-exponential decay of the signal increasing the b-values suggesting the presence of two different type of diffusion, called fast and slow.

Computer aided diagnosis within neuroimaging must rely on advanced image processing techniques to detect and quantify subtle signal changes that may be surrogate indicators of disease state. This thesis proposes two such novel methodologies that are both based on large volumes of interest, are data driven, and use cross-sectional scans: appearance-based classification (ABC) and voxel-based classification (VBC).
The concept of appearance in ABC represents the union of intensity and shape information extracted from magnetic resonance images (MRI). The classification method relies on a linear modeling of appearance features via principal components analysis, and comparison of the distribution of projection coordinates for the populations under study within a reference multidimensional appearance eigenspace. Classification is achieved using forward, stepwise linear discriminant analyses, in multiple cross-validated trials. In this work, the ABC methodology is shown to accurately lateralize the seizure focus in temporal lobe epilepsy (TLE), differentiate normal aging individuals from patients with either Alzheimer's dementia (AD) or Mild Cognitive Impairment (MCI), and finally predict the progression of MCI patients to AD. These applications demonstrated that the ABC technique is robust to different signal changes due to two distinct pathologies, to low resolution data and motion artifacts, and to possible differences inherent to multi-site acquisition.
The VBC technique relies on voxel-based morphometry to identify regions of grey and white matter concentration differences between co-registered cohorts of individuals, and then on linear modeling of variables extracted from these regions. Classification is achieved using linear discriminant analyses within a multivariate space composed of voxel-based morphometry measures related to grey and white matter concentration, along with clinical variables of interest. VBC is shown to increase the accuracy of prediction of one-year clinical status from three to four out of five TLE patients having undergone selective amygdalo-hippocampectomy. These two techniques are shown to have the necessary potential to solve current problems in neurological research, assist clinical physicians with their decision-making process and influence positively patient management.

This thesis describes a new melanoma diagnosii:s framework based on photometric stereo - a linear and computationally efficient method for extracting 3D data. The primary goal of this thesis is to investigate texture analysis of skin, including 3D based techniques, to realize a relatively reliable and accurate diagnosis of melanoma. In this thesis, a novel semi-automatic segmentation algorithm based on the normalized cut is proposed at first to realize fast segmentatioID. of melanoma images. As classic normalized cut involves generalized eigenvalue decompositioru, it can be extremely slow in segmenting large skin images even if the results are usually very impressive. Here a hierarchical normalized cut is proposed to conduct fast segmentation of skim lesion images with final results as good as the classic normalized cut in most cases. In addition, a new border analysis framework called centroid distance diagram is formulated to describe the border irregularity of melanoma images. This Fourier transform based approach gives a series of border irregularity descriptors insensitive to tire scate and rotation of tire images. Descriptors generated by using this method are proved to be effective features in describing malignant melanomas. The relationship between classic convexity and centroid convexity of border curves is examined in detail and the non-centroid -convexity index is generated to measure the irregularity of border curves. Moreover, statistical principal curvature patterns of skin surfaces are formulated to describe 3D melanoma shapes. The principal curvatures of skin surfaces are extracted by using the normal vector data obtained from photometric stereo. A robust estimator of these curvature parameters is embedded in this algorithm by using Gaussian kernels. Finally, an ensemble classifier for melanoma diagnosis is formulated by combining classifiers designed with various 2D /3D descriptors, e.g., border irregularity descriptors, colour variation descriptors and 3D principal curvature pattern descriptors. Even when the performance of the classifier obtained from one single group of descriptors is relatively poor, the ensemble classifier can offer highly impressive performances. Two novel schemes for designing this ensemble I classifier are proposed and compared with other well-known ensemble classifier designing methods, including Boosting and majority voting. Experiment studies suggest that the method named Bayesian II gives the best mean sensitivity (91.08 percent) and the best consistency level of sensitivity while the mean specificity can achieve 90.89 percent. The present work thus makes a novel contribution to the existing methods for computer aided diagnosis of melanoma. However, there are still many intriguing directions for further research works such as enhancing the accuracy of 3D data, recognition of subtypes of melanoma, introducing other descriptors in diagnosis and so on.

Le cancer du sein est un des plus répandus chez la femme. Le dépistage systématique permet de baisser le taux de mortalité mais crée une charge de travail importante pour les professionnels de santé. Des outils d’aide au diagnostic sont conçus pour réduire ladite charge, mais un niveau de performance élevé est attendu. Les techniques d’apprentissage profond peuvent palier les limitations des algorithmes de traitement d’image traditionnel et apporter une véritable aide à la décision. Néanmoins, plusieurs verrous technologiques sont associés à l’apprentissage profond appliqué à l’imagerie du sein, tels que l’hétérogénéité et le déséquilibre de données, le manque d’annotations, ainsi que la haute résolution d’imagerie. Confrontés auxdits verrous, nous abordons la problématique d’aide au diagnostic de plusieurs angles et nous proposons plusieurs méthodes constituant un outil complet. Ainsi, nous proposons deux méthodes d’évaluation de densité du sein étant un des facteur de risque, une méthode de détection d’anormalités, une technique d’estimation d’incertitude d’un classifieur basé sur des réseaux neuronaux, et une méthode de transfert de connaissances depuis mammographie 2D vers l’imagerie de tomosynthèse. Nos méthodes contribuent notamment à l’état de l’art des méthodes d’apprentissage faible et ouvrent des nouvelles voies de recherche
Breast cancer has the highest incidence amongst women. Regular screening allows to reduce the mortality rate, but creates a heavy workload for clinicians. To reduce it, the computer-aided diagnosis tools are designed, but a high level of performances is expected. Deep learning techniques have a potential to overcome the limitations of the traditional image processing algorithms. Although several challenges come with the deep learning applied to breast imaging, including heterogeneous and unbalanced data, limited amount of annotations, and high resolution. Facing these challenges, we approach the problem from multiple angles and propose several methods integrated in complete solution. Hence, we propose two methods for the assessment of the breast density as one of the cancer development risk factors, a method for abnormality detection, a method for uncertainty estimation of a classifier, and a method of transfer knowledge from mammography to tomosynthesis. Our methods contribute to the state of the art of weakly supervised learning and open new paths for further research

Includes bibliography.
With the improvement in computer technology, Computer Aided Diagnosis (CAD) is becoming an increasingly more powerful tool for radiologists. The focus of this project was on CAD of pulmonary miliary tuberculosis. Several methods for enhancing lung textures were discussed as an aid to the radiologist in diagnosing miliary TB. Some statistical approaches and template matching methods were used to measure characteristics of both healthy and unhealthy (miliary TB) lung textures. These measurements were evaluated to see if a computer can be programmed to differentiate between lung texture from a healthy lung and lung texture from a lung with miliary TB.

Parkinson‘s disease is a disorder of nervous system that mainly affects aged people. It is caused by the progressive degeneration of nerve cells in the brain, which results in a lack of dopamine necessary for controlled movements. As Parkinson‘s disease can not be cured, the early detection of Parkinson‘s disease is very helpful for a better management and care of patients, their families and the communities. This thesis reviewed the research status of Parkinson’s Disease (PD) diagnosis and extensive genetic association of PD issues, and proposed a key feature based machine learning strategy for PD classification tasks based on the open source database PPMI. In this thesis, a series of classification experiments were carried out, such that the set of single modality models were compared to a multiple modality model for PD diagnosis. Results of these experiments lead to a critical issue that more features for a diagnostic system seem to be necessary, but it should be verified if the more features are, the better the performance of a diagnostic system is. Therefore, a key feature classification strategy was proposed to explore the issue proposed above. At the first step, diagnostic features for PD classification are ranked by a decision tree model. This ranking procedure is then followed by a feeding procedure, in which a set of selected key features according to their importance ranking are fed into a machine learning model progressively, each time by feeding one more feature. This feeding procedure is iteratively processed until the performance reaches a local or global optimum. Based on the key feature classification strategy, an extensive investigation was carried out for the purpose to detect genetic association of PD. Results indicate that the proposed key feature based methodology leads to a more effective classification by a variety of machine learning models for PD diagnosis. The results presented in this thesis show that it is possible to conduct fewer key clinical examinations for PD diagnosis, or several conventional clinical examinations without expensive genetic sequencing studies to detect genetic associations to PD from other PD related categories successfully. The findings in this thesis have both social and economical values.

Medical imaging has advanced at a tremendous rate since x-rays were discovered in 1895. Today, x-ray machines produce extremely high-quality images for radiologists to interpret. However, the methods of interpretation have only recently begun to be augmented by advances in computer technology. Computer aided diagnosis (CAD) systems that guide healthcare professionals to making the correct diagnosis are slowly becoming more prevalent throughout the medical field. Bone fractures are a relatively common occurrence. In most developed countries the number of fractures associated with age-related bone loss is increasing rapidly. Regardless of the treating physician's level of experience, accurate detection and evaluation of musculoskeletal trauma is often problematic. Each year, the presence of many fractures is missed during x-ray diagnosis. For a trauma patient, a mis-diagnosis can lead to ineffective patient management, increased dissatisfaction, and expensive litigation. As a result, detection of long-bone fractures is an important orthopaedic and radiologic problem, and it is proposed that a novel CAD system could help lower the miss rate. This thesis examines the development of such a system, for the detection of long-bone fractures. A number of image processing software algorithms useful for automating the fracture detection process have been created. The first algorithm is a non-linear scale-space smoothing technique that allows edge information to be extracted from the x-ray image. The degree of smoothing is controlled by the scale parameter, and allows the amount of image detail that should be retained to be adjusted for each stage of the analysis. The result is demonstrated to be superior to the Canny edge detection algorithm. The second utilises the edge information to determine a set of parameters that approximate the shaft of the long-bone. This is achieved using a modified Hough Transform, and specially designed peak and line endpoint detectors. The third stage uses the shaft approximation data to locate the bone centre-lines and then perform diaphysis segmentation to separate the diaphysis from the epiphyses. Two segmentation algorithms are presented and one is shown to not only produce better results, but also be suitable for application to all long-bone images. The final stage applies a gradient based fracture detection algorithm to the segmented regions. This algorithm utilises a tool called the gradient composite measure to identify abnormal regions, including fractures, within the image. These regions are then identified and highlighted if they are deemed to be part of a fracture. A database of fracture images from trauma patients was collected from the emergency department at the Flinders Medical Centre. From this complete set of images, a development set and test set were created. Experiments on the test set show that diaphysis segmentation and fracture detection are both performed with an accuracy of 83%. Therefore these tools can consistently identify the boundaries between the bone segments, and then accurately highlight midshaft long-bone fractures within the marked diaphysis. Two of the algorithms---the non-linear smoothing and Hough Transform---are relatively slow to compute. Methods of decreasing the diagnosis time were investigated, and a set of parallelised algorithms were designed. These algorithms significantly reduced the total calculation time, making use of the algorithm much more feasible. The thesis concludes with an outline of future research and proposed techniques that---along with the methods and results presented---will improve CAD systems for fracture detection, resulting in more accurate diagnosis of fractures, and a reduction of the fracture miss rate.

The presence of distributed micro-calcifications can be an indicator of early breast cancer. On the mammogram, they appear as bright smooth particles superimposed on the normal breast image background. Radiologists determine the occurrence of this lesion by detecting the individual micro-calcifications and then examining their distribution within the breast tissue. Due to the visual complexity of the mammogram, the detection sensitivity is usually less than 100%. The digital environment has the potential to increase the radiologist's accuracy. We have developed a computer aided detection (CAD) scheme that can identify clinically indicative clusters of micro-calcifications. The CAD algorithm emulates some aspects of the radiologists' approach by using contrast texture energy segmentation and morphological distribution analysis. On a local database of 61 mammograms digitised at 100μm with 8 bits intensity resolution, the CAD returns: a) 85% sensitivity (91% for malignant lesions and 78% for those that are benign), b) 0.33 false positive clusters (FPC) per image and c) 92% specificity. Therefore, the output from the CAD is shown to compare favourably with the performance of an expert radiologist. It also compares favourably with other CAD techniques, exceeding many algorithms which employ a higher level of mathematical complexity. The scheme is tested on an international database provided by the Mammographic Image Analysis Society. In this case it returns a) 96.4% sensitivity (100% for malignant lesions and 92% for those that are benign) b) 2.35 FPC rate per image and c) 33% specificity. The higher FPC rate is attributed to the different acquisition and production of the digital mammograms. It is concluded that this can be reduced by employing a shape analysis procedure to the CAD's final output. It is shown that the image processing principles we have implemented are generally successful on databases which are produced at other centres under different technical conditions.

Various types of vascular diseases such as carotid stenosis, aneurysms, Arterio-venous Malformations (AVM) and Subarachnoid Hemorrhage (SAH) caused by the rupture of an aneurysm are some of the common causes of stroke. The diagnosis and management of such vascular conditions presents a challenge. In this dissertation we present a vascular analysis framework for Computer Aided Diagnosis (CAD) of the neuro-vasculature. We develop methods for 3D vascular decomposition, vascular skeleton extraction and identification of vascular structures such as aneurysms.

Owing to the complex and highly tortuous nature of the vasculature, analysis is often only attempted on a subset of the vessel network. In our framework we first, compute the decomposition of the vascular tree into meaningful sub-components. A novel spectral segmentation approach is presented that focuses on the eigenfunctions of the Laplace-Beltrami operator (LBO), FEM discretization. In this approach, we attain a set of vessel segmentations induced by the nodal sets of the LBO. This discretization produces a family of real valued functions, which provide interesting insights in the structure and morphology of the vasculature. Next, a novel Weighted Approximate Convex Decomposition (WACD) strategy is proposed to understand the nature of complex vessel structures. We start by addressing this problem of vascular decomposition as a cluster optimization problem and introduce a methodology for compact geometric decomposition. These decomposed vessel structures are then grouped into anatomically relevant sections using a novel vessel skeleton extraction methodology that utilizes a Laplace based operator. Vascular analysis is performed by obtaining a surface mapping between decomposed vessel sections. A non-rigid correspondence between vessel surfaces are achieved using Thin Plate Splines (TPS), and changes between corresponding surface morphologies are detected using Gaussian curvature maps and mean curvature maps. Finally, characteristic vascular structures such as vessel bifurcations and aneurysms are identified using a Support Vector Machine (SVM) on the most relevant eigenvalues, obtained through feature selection.

The proposed CAD framework was validated using pre-segmented sections of vasculature archived for 98 aneurysms in 112 patients. We first test our methodologies for vascular segmentation and next for detection. Our vascular segmentation approaches produced promising results, 81% of the vessel sections correctly segmented. For vascular classification, Recursive Feature Elimination (RFE) was performed to find the most compact and informative set of features. We showed that the selected sub-set of eigenvalues produces minimum error and improved classifier precision. This analysis framework was also tested on longitudinal cases of patients having internal cerebral aneurysms. Volumetric and surface area comparisons were made by establishing a correspondence between segmented vascular sections. Our results suggest that the CAD framework was able to decompose, classify and detect changes in aneurysm volumes and surface areas close to that segmented by an expert.

Hip replacement surgery has experienced a dramatic evolution in recent years supported by the latest developments in many areas of technology and surgical procedures. Unfortunately complications that follow hip replacement surgery remains the most challenging dilemma faced both by the patients and medical experts. The thesis presents a novel approach to segment the prosthesis of a THR surgical process by using an Active Contour Model (ACM) that is initiated via an automatically detected seed point within the enarthrosis region of the prosthesis. The circular area is detected via the use of a Fast, Randomized Circle Detection Algorithm. Experimental results are provided to compare the performance of the proposed ACM based approach to popular thresholding based approaches. Further an approach to automatically detect the Obturator Foramen using an ACM approach is also presented. Based on analysis of how medical experts carry out the detection of loosening and subsidence of a prosthesis and the presence of infections around the prosthesis area, this thesis presents novel computational analysis concepts to identify the key feature points of the prosthesis that are required to detect all of the above three types of complications. Initially key points along the prosthesis boundary are determined by measuring the curvature on the surface of the prosthesis. By traversing the edge pixels, starting from one end of the boundary of a detected prosthesis, the curvature values are determined and effectively used to determine key points of the prosthesis surface and their relative positioning. After the key-points are detected, pixel value gradients across the boundary of the prosthesis are determined along the boundary of the prosthesis to determine the presence of subsidence, loosening and infections. Experimental results and analysis are presented to show that the presence of subsidence is determined by the identification of dark pixels around the convex bend closest to the stem area of the prosthesis and away from it. The presence of loosening is determined by the additional presence of dark regions just outside the two straight line edges of the stem area of the prosthesis. The presence of infections is represented by the determination of dark areas around the tip of the stem of the prosthesis. All three complications are thus determined by a single process where the detailed analysis defer. The experimental results presented show the effectiveness of all proposed approaches which are also compared and validated against the ground truth recorded manually with expert user input.

Inflammatory muscle disease (myositis) is characterised by inflammation and a gradual increase in muscle weakness. Diagnosis typically requires a range of clinical tests, including magnetic resonance imaging of the thigh muscles to assess the disease severity. In the past, this has been measured by manually counting the number of muscles affected. In this work, a computer-aided analysis of inflammatory muscle disease is presented to help doctors diagnose and monitor the disease. Methods to quantify the level of oedema and fat infiltration from magnetic resonance scans are proposed and the disease quantities determined are shown to have positive correlation against expert medical opinion. The methods have been designed and tested on a database of clinically acquired T1 and STIR sequences, and are proven to be robust despite suboptimal image quality. General background information is first introduced, giving an overview of the medical, technical, and theoretical topics necessary to understand the problem domain. Next, a detailed introduction to the physics of magnetic resonance imaging is given. A review of important literature from similar and related domains is presented, with valuable insights that are utilised at a later stage. Scans are carefully pre-processed to bring all slices in to a common frame of reference and the methods to quantify the level of oedema and fat infiltration are defined and shown to have good positive correlation with expert medical opinion. A number of validation tests are performed with re-scanned subjects to indicate the level of repeatability. The disease quantities, together with statistical features from the T1-STIR joint histogram, are used for automatic classification of the disease severity. Automatic classification is shown to be successful on out of sample data for both the oedema and fat infiltration problems.

Breast cancer is now one of the most common forms of cancer and the leading cause of mortality in women in the developed countries. Early detection of breast cancer is currently the way to reduce breast cancer mortality and enhance the cure rate. Mammogram screening is widely recognized as the most reliable method for early detection of lesions and clustered microcalcifications, which are the two prominent symptoms of breast cancer. This thesis presents an image processing procedure for the automatic detection of clustered microcalcifications in digitized mammograms. This method consists of two main steps. First, possible microcalcification pixels in the mammograms are segmented out using wavelet features, and grouped into potential individual microcalcification objects by their spatial connectivity. Second, individual microcalcifications are detected by using the structure features extracted from the potential microcalcification objects. The classifiers used in the two steps are feedforward neural networks. The method is applied to 40 regions of interest extracted from mammograms in the Nijimegen database containing 144 clusters of microcalcifications. Results show that the proposed procedure gives satisfactory detection performance. In particular, a 97 percent mean true positive detection rate is achieved at the cost of 1.67 false positive in the whole dataset.

Environ 150.000 personnes souffrent en France d'une épilepsie partielle réfractaire à tous les médicaments. La chirurgie, qui constitue aujourd’hui le meilleur recours thérapeutique nécessite un bilan préopératoire complexe. L'analyse de données d'imagerie telles que l’imagerie par résonance magnétique (IRM) anatomique et la tomographie d’émission de positons (TEP) au FDG (fluorodéoxyglucose) tend à prendre une place croissante dans ce protocole, et pourrait à terme limiter de recourir à l’électroencéphalographie intracérébrale (SEEG), procédure très invasive mais qui constitue encore la technique de référence. Pour assister les cliniciens dans leur tâche diagnostique, nous avons développé un système d'aide au diagnostic (CAD) reposant sur l'analyse multivariée de données d'imagerie. Compte tenu de la difficulté relative à la constitution de bases de données annotées et équilibrées entre classes, notre première contribution a été de placer l'étude dans le cadre méthodologique de la détection du changement. L'algorithme du séparateur à vaste marge adapté à ce cadre là (OC-SVM) a été utilisé pour apprendre, à partir de cartes multi-paramétriques extraites d'IRM T1 de sujets normaux, un modèle prédictif caractérisant la normalité à l'échelle du voxel. Le modèle permet ensuite de faire ressortir, dans les images de patients, les zones cérébrales suspectes s'écartant de cette normalité. Les performances du système ont été évaluées sur des lésions simulées ainsi que sur une base de données de patients. Trois extensions ont ensuite été proposées. D'abord un nouveau schéma de détection plus robuste à la présence de bruit d'étiquetage dans la base de données d'apprentissage. Ensuite, une stratégie de fusion optimale permettant la combinaison de plusieurs classifieurs OC-SVM associés chacun à une séquence IRM. Enfin, une généralisation de l'algorithme de détection d'anomalies permettant la conversion de la sortie du CAD en probabilité, offrant ainsi une meilleure interprétation de la sortie du système et son intégration dans le bilan pré-opératoire global
One third of patients suffering from epilepsy are resistant to medication. For these patients, surgical removal of the epileptogenic zone offers the possibility of a cure. Surgery success relies heavily on the accurate localization of the epileptogenic zone. The analysis of neuroimaging data such as magnetic resonance imaging (MRI) and positron emission tomography (PET) is increasingly used in the pre-surgical work-up of patients and may offer an alternative to the invasive reference of Stereo-electro-encephalo -graphy (SEEG) monitoring. To assist clinicians in screening these lesions, we developed a computer aided diagnosis system (CAD) based on a multivariate data analysis approach. Our first contribution was to formulate the problem of epileptogenic lesion detection as an outlier detection problem. The main motivation for this formulation was to avoid the dependence on labelled data and the class imbalance inherent to this detection task. The proposed system builds upon the one class support vector machines (OC-SVM) classifier. OC-SVM was trained using features extracted from MRI scans of healthy control subjects, allowing a voxelwise assessment of the deviation of a test subject pattern from the learned patterns. System performance was evaluated using realistic simulations of challenging detection tasks as well as clinical data of patients with intractable epilepsy. The outlier detection framework was further extended to take into account the specificities of neuroimaging data and the detection task at hand. We first proposed a reformulation of the support vector data description (SVDD) method to deal with the presence of uncertain observations in the training data. Second, to handle the multi-parametric nature of neuroimaging data, we proposed an optimal fusion approach for combining multiple base one-class classifiers. Finally, to help with score interpretation, threshold selection and score combination, we proposed to transform the score outputs of the outlier detection algorithm into well calibrated probabilities

Alzheimer’s disease (AD) is a chronic and progressive, irreversible syndrome that deteriorates the cognitive functions. Official death certificates of 2013 reported 84,767 deaths from Alzheimer’s disease, making it the 6th leading cause of death in the United States. The rate of AD is estimated to double by 2050. The neurodegeneration of AD occurs decades before symptoms of dementia are evident. Therefore, having an efficient methodology for the early and proper diagnosis can lead to more effective treatments. Neuroimaging techniques such as magnetic resonance imaging (MRI) can detect changes in the brain of living subjects. Moreover, medical imaging techniques are the best diagnostic tools to determine brain atrophies; however, a significant limitation is the level of training, methodology, and experience of the diagnostician. Thus, Computer aided diagnosis (CAD) systems are part of a promising tool to help improve the diagnostic outcomes. No publications addressing the use of Feedforward Artificial Neural Networks (ANN), and MRI image attributes for the classification of AD were found. Consequently, the focus of this study is to investigate if the use of MRI images, specifically texture and frequency attributes along with a feedforward ANN model, can lead to the classification of individuals with AD. Moreover, this study compared the use of a single view versus a multi-view of MRI images and their performance. The frequency, texture, and MRI views in combination with the feedforward artificial neural network were tested to determine if they were comparable to the clinician’s performance. The clinician’s performances used were 78 percent accuracy, 87 percent sensitivity, 71 percent specificity, and 78 percent precision from a study with 1,073 individuals. The study found that the use of the Discrete Wavelet Transform (DWT) and Fourier Transform (FT) low frequency give comparable results to the clinicians; however, the FT outperformed the clinicians with an accuracy of 85 percent, precision of 87 percent, sensitivity of 90 percent and specificity of 75 percent. In the case of texture, a single texture feature, and the combination of two or more features gave results comparable to the clinicians. However, the Gray level co-occurrence matrix (GLCOM), which is the combination of texture features, was the highest performing texture method with 82 percent accuracy, 86 percent sensitivity, 76 percent specificity, and 86 percent precision. Combination CII (energy and entropy) outperformed all other combinations with 78 percent accuracy, 88 percent sensitivity, 72 percent specificity, and 78 percent precision. Additionally, a combination of views can increase performance for certain texture attributes; however, the axial view outperformed the sagittal and coronal views in the case of frequency attributes. In conclusion, this study found that both texture and frequency characteristics in combinations with a feedforward backpropagation neural network can perform at the level of the clinician and even higher depending on the attribute and the view or combination of views used.

Over the years, exemplar-based methods have yielded significant improvements over their model-based counterparts in image synthesis applications. Notably, texture synthesis algorithms using an exemplar-based approach have shown success where traditional stochastic methods failed. As an illustrative example, I will present an exemplar-based approach that yields substantial benefits for user-guided terrain synthesis using Digital Elevation Models (DEMs). This success is realized through exploitation of structural properties of natural terrain. In addition to their proliferation in the image synthesis domain, as annotated image datasets become increasingly available, exemplar-based methods are also gaining in popularity for image analysis applications. This thesis addresses the intersection between exemplar-based analysis and the problem of content-based image retrieval (CBIR). A basic problem in CBIR is the process by which the search criteria are refined by the user through the manipulation of returned exemplars. Exemplar-based analysis is particularly well-suited to query refinement due to its interpretability and the ease with which it can be incorporated into an interactive system. I investigate this connection in the domain of Computer-Assisted Diagnosis (CAD) of dermatological images. I demonstrate that exemplar-based approaches in CBIR can be effective for diagnosing pigmented skin lesions (PSLs). I will present an exemplar-based algorithm for segmenting PSLs in dermatoscopic images. In addition, I will present a generalized representation of dermoscopic features for detection and matching. This representation not only leads to an exemplar-based PSL diagnosis scheme, but it also enables us to realize interactive region-of-interest retrieval, which includes a relevance feedback mechanism to facilitate more flexible query-by-example analysis. Finally, I will assess the benefit of this CBIR-CAD approach through both quantitative evaluations and user studies.

博士
國立中山大學
資訊工程學系研究所
94
As computer technologies are developed rapidly in recent years, the ways to diagnose diseases also alter in clinical practice. Picture Archiving and Communication System (PACS) is an example that makes the diagnostic way for medical images change from view box to monitor. All types of medical images tend to be digitized and this makes it practical for helping doctor diagnose medical images via computer technologies. In this thesis, we propose a systemic approach to screen abnormalities in chest radiographs. First, a preprocess step identifying the view of chest radiographs is introduced. Second, a method is proposed for automated detection of gross abnormal opacity in chest radiographs. Third, computation time reduction is performed by subsampling. Finally, a computer-aided diagnosis system is implemented and evaluated in a clinical environment. Major technique used in this thesis is to analyze the projection profile obtained by projecting a chest image on to the mediolateral axis. The discriminant performance for each method is evaluated by using receiver operating characteristic (ROC) analysis. The results indicate that the proposed methods are potentially useful for screening abnormalities in chest radiographs.

碩士
長庚大學
電機工程研究所
86
The related research of the Chinese Medical Diagnosis has been widely studied and discussed. Many people desire to construct a specific and objective model of traditional Chinese Medical Diagnosis process. To modernize the Chinese Medical System with scientific thinking and methodology and how to reserve the original theoretical characteristic become an important task. In this thesis, we simulate the diagnosis of Chinese Medicine for asthma with fuzzy set theory. Asthma present many inherent or extrinsic symptom of human body. Each symptom is cured with different correspondent Chinese Medicine. Because of the symptom provided by the clinical patients is almost indistinct or incomplete, scientific diagnosis Chinese Medicine with these symptoms is difficult. Experienced Chinese Medical doctors infer the disease based on some cricerions and combine the condition and history of disease to complete the diagnosis. In this study, we construct a computerized Chinese Medical diagnosis system with fuzzy set theory. The source of data are obtained via looking, listening, asking and touching. These data may be incomplete or erroneouss but the fuzzy set theory has the ability to deal with this kind of situation. We take use of the property of fuzzy set theory to extract the relationship between asthma and its symptoms. The final object is to simulate the diagnose process of Chinese Medical doctors for asthma and buildup the Fuzzy Expert Chinese Medical diagnose System. We merge the fuzzy set theory and Chinese Medical Diagnose to construct a computerized Chinese Medical Diagnose System with capability of incomplete and erroneous symptom data. The fuzzy set theory has been widely evolved in many applications include the medical use. The traditional logic theory is binary with true value ?" and false value ?". There is no so called "fuzzy zone". But the uncertainty often encountered and can not be completely described with traditional logic set theory. The fuzzy set logic which is exceeded from the traditional set logic include arbitrary value between *0* and *1*. It represent the logic of human beens and Chinese Medical Diagnosis. It extract the relationship of symptom-to- symptom, symptom-to-syndrom and syndrom-to- syndrom with fuzzy relation matrix. This will fully represent the global concept of Chinese Medical Diagnosis. In the future, the system parameter can be adjust to be utilized in diagnosis of other symptom with more clinical data. If support the property diagnosis and alert of possible diseases.

博士
國立臺灣大學
臨床醫學研究所
102
Thyroid cancer is the most common endocrine malignancy. The incidence is increasing in the past decades. How to improve the detection rate and diagnostic accuracy has become an urgent issue. We proposed the novel computerized methods in this thesis to quantify the ultrasonic features of thyroid cancer. We included two ultrasonic features which were microcalcifications and heterogeneity in this thesis to test whether computerized method could be helpful in diagnosis of thyroid cancer. The first part was to improve the ultrasonographic detection rates of thyroid cancers with microcalcifications, we proposed to enhance the sensitivity of sonographic calcifications detection and to avoid inter-observer variation by a computerized quantification method in a prospective setting. A total of 256 nodules were included (173 benign, 83 malignant). Among them, the diagnosis of 181 nodules (102 benign, 79 malignant) were verified by surgical pathology. Quantification of cystic components and calcifications was automatically performed by a proprietary program (AmCAD-UT) implemented with methods proposed in this paper. The calcification index (CI) between benign and malignant nodules diagnosed by combined FNA biopsy and surgical pathology results (total number, 256) showed a significant difference (p<0.001, AUC=0.746). Furthermore, we excluded patients without surgical pathology results for further validation, and the CI between benign and malignant nodules confirmed by pathology results (total number, 181) showed a significant difference (p<0.001, AUC=0.763). To learn whether our computer program increased our diagnostic capabilities, we analyzed human investigators and their abilities to detect and evaluate. In this study, calcifications were noted in 48.19% (40 of 83) of malignant thyroid nodules and in 10.98% (19 of 173) of benign nodules. This new computer-aided diagnosis method to evaluate the sonographic calcifications of thyroid nodules is a more sensitive and more objective method. It can provide better sensitivity than conventional methods in the diagnosis of thyroid malignancies containing microcalcifications. The second part is to test whether the computerized quantification of ultrasonic heterogeneity can aid the diagnosis of thyroid malignancy, we evaluated ultrasonic heterogeneity by an objective and quantitative computerized method in a prospective setting. A total of 400 nodules including 271 benign thyroid nodules and 129 malignant thyroid nodules were evaluated. Quantification of ultrasonic heterogeneity (heterogeneity index, HI) was performed by a proprietary program implemented with methods proposed in this paper. The HI values between benign and malignant nodules, diagnosed by a combination of fine-needle aspiration (FNA) and surgical pathology results, are significantly different (p<0.001, AUC=0.714). Ultrasonic heterogeneity (US-H) of these samples assessed by an experienced clinician could not significantly differentiate between benign and malignant thyroid nodules. However, the nodules with marked heterogeneity of US-H showed higher HI values than that of nodules with homogeneous US-H assessment. These results indicate that using the new computer-aided diagnosis method to evaluate the ultrasonic heterogeneity of thyroid nodules is an objective and quantitative method that is correlated to the conventional US-H assessment but can aid in the diagnosis of thyroid malignancy better.

碩士
元智大學
通訊工程學系
94
Traditionally, diagnosis of liver disease such as liver cyst, cavernous hemangioma and hepatoma is heavily dependent on professional radiologists. However, it is not guaranteed to make a highly accurate decision even for a specialist with a lot of experience. We propose a scheme of computer-aided diagnosis system which aims to assist radiologists in making more precise diagnosis of liver diseases. First of all, we select the region of interests (ROIs) from images with appropriate size. Secondly, the features including gray-level, co-occurrence matrix, and a bank of Gabor filters are extracted from the selected ROIs. Then both the methods of principle component analysis (PCA) and sequential forward selection (SFS) are used to reduce the dimension of feature vectors. The reduced features are fed into classifiers in which support vector machine (SVM) and radial basis neural network (RBFNN) are employed as the techniques for the classifiers. Finally, the analysis of receiver operating characteristic (ROC) curve is implemented to evaluate the performance of this system for the sake of getting higher distinction efficiency.

碩士
國立臺灣大學
醫學工程學研究所
89
To early detect the malignant breast lesions and to reduce the number of unnecessary biopsies, computer-aided diagnosis (CAD) of sonographic breast lesions have been studied extensively. Although reasonable performances have been achieved, the conventional CAD algorithms generally suffer the unreliable features and the inability to utilize the insignificant features, which may be decisive in determining the equivocal cases for the significant features. To overcome these two problems, in this paper, we proposed a set of highly effective geometrical features and a new two-level classification scheme, which took into account the value of the insignificant features. The first-level classification was composed of a logistic discrimination function for selecting the significant features and a multi-layer feed-forward neural network for classification. The second-level classification has been formulated as a constrained maximization problem designed to fully utilize the insignificant features. The experimental results show that the best classification accuracy, sensitivity, specificity, positive prediction value and negative prediction value for the proposed two-level classification scheme with 160 lesions were as high as 93.8%, 91.3%, 95.6%, 93.5%, and 94.0%, respectively. By considering the insignificant features, we have succeeded in increasing the classification accuracy, the sensitivity and the specificity by 3.2%, 4.3%, and 2.2%, respectively.

Computer aided diagnosis is a relatively new field, through the use of new techniques algorithms and technologies, it can help technicians perform a better and faster analysis, reduce or even substitute part of their workload. Patents are windows into a company's technological assets, as well as into the state of a certain technology field. In this thesis we analyzed patents that are mainly related to the automated analysis of human retinopathies. Using patent search engines we explored the various patent databases, using keywords related to the area and the international patent classification to refine the search and eliminate unrelated results, proceeding then to a thorough analysis of the dataset. By analyzing the structured and unstructured text, contained in the obtained patents, different observations where made: major players in the field,patent timelines, main technologies involved and the direction of the technology evolution.

碩士
中華大學
生物資訊學系碩士班
100
With the advances in mass communication, there are many ways that people can acquire knowledge. Because of the fast pace of living, people are suffering from various diseases, especially chronicle diseases. Now a day, patients usually choose western medicine, because it can reduce pain in a short time. By comparison, most of people have the impression that it takes a longer time to improve their health condition using Chinese medicine. Because Western medicine has been less effective in the treatment of chronic diseases, traditional Chinese medicine is gradually getting more attention. In addition to herbal medicines, Chinese medicine also, acupuncture, meridian massage, and healthy diets to improve personal health condition. People must to make an appointment at hospital when they are sick. If we can preliminary diagnose by myself. This will add considerable convenience, and even make Chinese medicine practitioners to increase the convenience of the diagnosis. The most important thing is to help users to prevent diseases by traditional Chinese medicine health care knowledge of the system. Therefore, they need a simple and containing a wealth of information diagnosis system for Chinese medicine. This system uses information of user’s symptoms to match and calculate the results of the highest probability of diagnosis by SQL Server's "SELECT" function. According to the results of diagnosis, the system provides knowledge of Chinese herbal medicine and Chinese meridian-point conditioning. Finally, this research uses a virtual example to illustrate possible applications of the system.

Computer aided diagnosis is a relatively new field, through the use of new techniques algorithms and technologies, it can help technicians perform a better and faster analysis, reduce or even substitute part of their workload. Patents are windows into a company's technological assets, as well as into the state of a certain technology field. In this thesis we analyzed patents that are mainly related to the automated analysis of human retinopathies. Using patent search engines we explored the various patent databases, using keywords related to the area and the international patent classification to refine the search and eliminate unrelated results, proceeding then to a thorough analysis of the dataset. By analyzing the structured and unstructured text, contained in the obtained patents, different observations where made: major players in the field,patent timelines, main technologies involved and the direction of the technology evolution.

碩士
國立臺灣大學
資訊工程學研究所
106
Among female throughout the world, breast cancer has become one of the most common carcinomas and the leading cause of cancer-related death. Early detection can provide a better treatment and significantly reduce mortality. Currently, the most effective tool to diagnose breast cancer is mammography screening. Tomosynthesis as a three dimensional (3-D) tomographic technique can overcome the overlapping problem from superimposed tissues of two dimensional (2-D) mammography. Therefore, we proposed a computer-aided diagnosis (CADx) system implemented in tomosynthesis and also in mammography to compare their performance. The CADx system was built by binary logistic regression classifier. Texture features, including gray-level co-occurrence matrix (GLCM), ranklet, and Gabor, were extracted from user-specified regions of interest (ROIs) in mammograms or volumes of interest (VOIs) in tomosynthesis images. The performance of different combinations of features were evaluated. The CADx system was tested with a dataset of 42 benign and 82 malignant tumors. The best performance was achieved by applying Gabor feature in tomosynthesis with an accuracy of 85.48% (106/124), a sensitivity of 86.59% (71/82), a specificity of 83.33% (35/42), and an Az value of 0.8712. To summarize, tomosynthesis is more effective in classification of breast tumor with Gabor feature than mammography.

博士
臺灣大學
醫學工程學研究所
98
Intracranial hematomas, either traumatic or spontaneous, can produce fatal outcomes because they can produce local pressure on the brain. Accurate diagnosis and rapid decision making are the key factors to good patient outcome. This thesis introduces new methods capable of obtaining the features of the intracranial hematomas in brain CT images. In addition, a new approach of image segmentation integrating binary level set method and multi-resolution processing is proposed. We develop the decision rules to recognize the type of the intracranial hematoma on CT slices with large intracranial hematomas using C4.5 algorithm. These decision rules work well in different resolutions. To obtain robust segmentation of the intracranial hematoma regions, we introduce a multi-resolution binary level set method using image pyramids and apply it to hematoma segmentation. Prior to segmentation of the hematoma from the brain, anatomical knowledge is integrated with image processing techniques in the segmentation of intracranial regions. The results show excellent precision and recall as verified by human experts. In the second half of this thesis, we describe two methods for automatic measurement of the midline shift (MLS). The first one employs symmetry and curve fitting to measure the MLS of the CT slice at the level of Foramen of Monro. Genetic algorithm is used for parameter optimization. Landmark-based MLS recognition is carried out by first segmenting the frontal horn region followed by a knowledge-driven rule. Hough transform (HT) is then applied to locate the septum pellucidum. Finally, we describe automatic recognition of the basal cisterns using HT. This method is able to pick out the normal or compressed basal cistern region from the given CT data set.

碩士
中原大學
醫學工程學系
83
The research was to develop a computer-aided system for pulse diagnosis in Chinese medicine. The system is composed of five modes : 1) patient data management mode 2) pulse and ECG s- ignals acquisition mode 3) pulse signal analysis mode 4) pulse signal recognition mode 5) pulse diagnosis database. With this system the pulse diagnosis can be more scientific and objective than that with Chinese doctor''s finger tips. In the results, ma- inly pulse attribute that consists of "pulse shape" was In addition, pulse attributes consisting of "pulse strength" and "pulse position" and "pulse rate" and "pulse regularity", were directly estimated from the quantitative values which reach to certain levels. In the study, there was 16/18 correctness for "normal pulse" in pattern recognition, 7/8 correctness for "string pulse", 2/3 correctness for "rough pulse", 6/7 correctness for "sliding ", and 9/11 correctness for "string- rough pulse," the total ectness is about 80%. In the system, we can recognize "float pulse", "profundity ulse", "tardy pulse", "soon pulse", "scud pulse", rity pulse","slow-unreqularity pulse", "weak pulse", "soft , "rough pulse", "string pulse", "sliding pulse" and the complex pulses about them . In the system, some pulse attributes could not be that for the reasons as follows: (1).To get some special pulse attribute samples is not easy. (2).Some pulse attributes have not standard rules to recognize. (3).Some pulse attributes must detect with three finger tips,but the system hardware uses one finger tip mold( one sensor) to Thereafter, the collections of pulse waves and the definitions pulse recognizable standards will be the most work in the future research. If the pulse diagnosis was precise definition as ECG the auto recognition of pulse wave will be more efficient and

碩士
中原大學
醫學工程研究所
94
Abstract Cervical cancer is one of deadly diseases of cancers with highest occurrence rate for women in Taiwan. Pap smear is the best inspection examination to prevent cervical cancer. In viewing the pap smear, normal cells and abnormal cells which including Low grade Squamous intraepithelial lesion (LSIL) and High grade Squamous intraepithelial lesion (HSIL) were distinguished by physician. The purpose of this study is applied the information technology to develop a system that can analyze the cell types and characteristic parameters of cervical cancer, and assist physicians to diagnosis cervix cancer in earlier stage. Some techniques such as: image processing, Back Propagation Networks (BPN) training, database storage, suitability assessment and analyzing, and interface development were used in this study. First of all, color cell images were transformed into gray level images. Through noises removed, morphology, and chain code techniques, the contour of image was circled and recorded. Then the gray and color cell images were segmented into three areas which are background, cytoplasm, and nuclear. Feathers that include RGB, HIS, Entropy, Contrast, and Nuclear/Cytoplasm (N/C) ratio were acquired from images, and the result from statistical analysis were served as distinguish parameters for the development of BPN. In order to train the BPN model, 120 image cases including 60 normal and 60 abnormal cases were used in this study. Then, 32 testing images including 11 normal and 21 abnormal cases and clinical diagnosis results were used to evaluate the accuracy of system. Moreover, according to the opinions from doctors, a friendly interface was developed for increasing the practicality of this system. Results show that P values from statistical analysis and the training result from BPN is highly correlated. For training images, the system perform well in distinguish cell type with accuracy, sensitivity, specificity were 1. For the testing images, the accuracy, sensitivity, specificity and kappa value were 0.97, 1, 0.91 and 0.93, respectively. Moreover, the N/C ratio for normal cells is less 0.1, for LSIL is between 0.1 and 0.2 , and for HSIL is between 0.2~1. Although the weighting value for N/C ratio is highest in BPN model, the color and other gray parameters are also play important roles to diagnose. Color, gray level, and the distribution of N/C ratio were used to distinguish normal and abnormal cells which include LSIL and HSIL. During this study, only one normal inflammatory cell case in test image was misjudged, because of some effects like noisy background, mucus and other complexities effects that can affect the results of diagnosis. In conclusion, compared with previous study this system is improved in efficient and accuracy. This system could not only segment and analyze multiple cells at one time, but also provide better result with new parameters selection. In the near future, we hope this system can provide more useful diagnosis information from cells for physicians, and can solve any relational problems about pap smear more efficiently thru database communication.

碩士
長庚大學
醫療機電工程研究所
93
The human brain degeneration cases have tremendously increased in the last decades. The major human brain degeneration diseases are Spinocerebellar Ataxia (SCA) and Multiple System Atrophy (MSA). The traditional clinical diagnosis on humans’ cerebellar atrophy is based on humans’ manual measurements by examining the difference of the measured amounts of the cerebellum and patients’ gene information. When the modern computer ability increases twice every 1.5 years, it’s a good approach using computer as a clinical diagnosis auxiliary to avoid humans’ manual errors. Based on this concept, a computer-aid diagnosis system of cerebellar degeneration diseases is proposed in this thesis. First, the differences of the measured areas of normal peoples’ and SCA and MSA patients’ MRI images are analyzed to build a feature database and distinguish between normal human, SCA patients and MSA patients. An enhanced segmentation method including Modified Fuzzy Connectedness Method and Gradient Vector Flow Active Contour Model is used to extract tissue areas with more accurate boundaries in this stage. Next, when a SCA or MSA patient is suggested in the previous stage, his gel electrophoresis images are analyzed to extract gene information for gene matching with a gene information database. These two stages work at the automatic or semi-automatic way with a friendly user interface to avoid humans’ operating errors. The establishment of this expert system with both feature and gene information will be helpful for physicians’ diagnosis and treatment for the SCA and MSA patients.

碩士
國立臺灣大學
資訊網路與多媒體研究所
104
In women aged 20-59 years, breast cancer is the highest mortality. Early treatment can effectively reduce the mortality of breast cancer. In recently, breast ultrasound image is often used to diagnose between benign and malignant tumors. For increasing the accuracy, most researches segment the tumor before classification, and the segmented results directly affect the classification between benign and malignant tumors. Therefore, the purpose of this study is to refine segmented tumors using the image matting method for computer-aided diagnosis. First, the volume-of-interest (VOI) of tumor was extracted from the ultrasound image and pre-segmented by a conventional segmentation method. The tri-map including the background, foreground, and unknown region was created with the pre-segmented tumor, and then the image matting method was applied for refining the segmentation according to the unknown region of tri-map. Texture and morphology features were extracted from refined segmentation result and then the support vector machine was applied with extracted features to classify tumor into benign or malignant tumor. This study was validated with 80 cases including 40 benign and 40 malignant breast lesions. According to the experiment results, applying the image matting method had better performance than not applying the image matting method, and the combination of GLCM, ranklet, and ellipsoid fitting feature set had significant (resolution??. The accuracy, sensitivity, specificity, and the area under ROC achieved 85.0% (68/80), 87.5% (35/40), 82.5% (33/40), and 0.8829, respectively. From the experiment results, the image matting method could actually refine tumor segmentation, and more precise classification between benign and malignant tumor results were obtained.

碩士
中原大學
醫學工程研究所
90
In radiology, it''s quite difficult to make a preliminary diagnosis of lung cancer by using a chest X-ray image. When an uncertainly abnormal tissue occurred in X-ray image, some advanced tests, such as CT, MRI, bronchotomy , or microtomy , will be made for more detailed diagnosis. The aim of this study was to develop an algorithm applying in detecting lung nodules on chest radiological image, and mark possible region of suspicious nodules for doctor''s diagnosis. The detecting method was based on difference image. At first, we selected a threshold for all image and mark all suspicious nodules regions by circularity at difference image. Then, two algorithms were used on the suspicious regions for reducing the number of false positive(FP). A biological information which is determine whether calcification by gray scale and verify whether mediastinum by position was used to reduce the number of FP. If it didn''t show great effect, artificial neural network (ANN) was applied to reduce FP number. The source of our image contained phantom and clinical chest X-ray image. By using of phantom image , the correctness of our algorithm could be evaluated. The algorithm was applied on real clinical patient image to evaluate its clinically practical. Besides, by the aids of clinical doctors, we found that the difference of performance between our system and traditional method could be distinguished. Our system provided the calculation of suspicious area and successfully detect the nodules in phantom. In real image, the sensitivity approaches to 100% for less nodules patients when applied the gray level of 30% in cumulative histogram as a threshold. With the aids of biological information and ANN, the FP number went down from 24.92/per image to 3.06/per image. It satisfied the demand of medical diagnosis. For more nodules image, the sensitivity approached to 0.96667 when applied 32% gray level as a threshold. With the aids of biological information and ANN, the FP number west down from 17.33/per image to 2.00/per image. Comparing with other systems, this system obtained less FP and higher accuracy. Result showed that we can detect nodules in either phantom or real image successfully with less FP number. Besides, our system provides flexible user''s platform. In the future, some functions, such as area and calculation time ,will be improved so that the system can be more clinically practical and achieve the goal of assisting the diagnosis of doctor.

博士
國立中正大學
資訊工程所
94
Ultrasound (US) is a valuable adjunct to mammography in breast imaging. Recently, relevant studies have been applied to computer-aided diagnosis (CAD) systems which have demonstrated a promising result for the classification of breast lesions. Although the conventional two-dimensional (2-D) US techniques of the breast at present have been widely used in surgical clinical practice, 2-D images are not enough to transmit the entire characteristics of a solid breast lesion. On the other hand, three-dimensional (3-D) US can fully provide the architecture of breast tumors in all aspects and therefore is capable of offering a more comprehensive way to characterize pathological features. The 3-D CAD systems are expected to be a useful CAD tool for classifying benign and malignant tumors in ultrasound, and can assist physicians to reduce misdiagnosis. In this study, we present our undergoing developments on 3-D breast ultrasound regarding two important properties that are frequently sighted in breast carcinoma, including spiculation, angiogenesis. Our work involves the techniques of how to extract features from the original 3-D images and of how to quantify them. The numerical data was finally related to tumor outcome and showed a good agreement with the delineation of the experts.

碩士
國立臺灣大學
資訊工程學研究所
99
The breast cancer is the most common cancer of women in the world and it is the major cause of the death for women in recent years. However, it is a type of cancer which has an excellent curability if it can be detected in early stage. Recently the computer-aided diagnosis systems can provide radiologists not only the existence information of the tumors but also the lesion classification of malignancy and benignancy. In this paper, the dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is used to detect changes of the contrast agent (CA) concentration over time which are modeled as a result of the exchange of CA molecules between compartments. Furthermore, the diffusion-weighted magnetic resonance imaging (DWI) is used to provide the information about local characteristic of water diffusion of biological tissues. An experienced radiologist indicates the tumor in DCE-MRI and a region growing based algorithm is then applied to segment the tumor. After the manual registration process between DCE-MRI and DWI by a radiologist, the tumor region in corresponding DCE-MRI is mapped into DWI. Then a modified fuzzy c-means clustering is used to identify the most possible malignant kinetic curve of the segmented tumor for analysis. The Tofts pharmacokinetic model is used to fit the kinetic curve of the tumor and the parameters of the model are used as the diagnosis features. Furthermore, the three-dimensional (3-D) morphology features, shape and texture, from DCE-MRI are proposed to improve the diagnosis performance. The shape features including compactness, margin, and ellipsoid fitting could describe the 3-D shape information of the tumor and the 3-D texture features based on the grey level co-occurrence matrix is also used to analyze the lesions. Besides, the apparent diffusion coefficient (ADC) features extracted from DWI give the quantitative measurement for the water diffusion of a lesion are also included in this study. In the experiments, 138 biopsy-proved lesions with 54 benign and 84 malignant used to evaluate the performance of the proposed diagnosis system for breast MRI. Its accuracy, sensitivity, specificity, and Az value are up to 91.30% (126/138), 92.86% (78/84), 88.89% (48/54), and 0.9333, respectively. From the experiment result, the conventional kinetic characteristics features, Tofts features and ADC features could have the better performance. Especially, the ADC features have the best sensitivity.

Οι μικροαποτιτανώσεις είναι από τις πιο σημαντικές ενδείξεις παθήσεων του μαστού και μπορεί να αποτελέσουν πρώιμη ένδειξη καρκίνου του μαστού. Πρόκειται για εναποθέσεις ασβεστίου στο μαστό με τη διάμετρός τους να κυμαίνεται από 0.1 έως 1 mm και εμφανίζονται είτε μόνες είτε σε ομάδες. Η ακριβής τμηματοποίηση (segmentation) των μικροαποτιτανώσεων στη μαστογραφία συνεισφέρει στην εξαγωγή αξιόπιστων χαρακτηριστικών μορφολογίας, που χρησιμοποιούνται στην αυτόματη διάγνωση με υπολογιστή (Computer-aided Diagnosis, CADx). Στα πλαίσια της παρούσης διδακτορικής διατριβής προτείνεται μία νέα μέθοδος τμηματοποίησης μικροαποτιτανώσεων, η οποία αρχικά εντοπίζει σημεία του περιγράμματος αυτών. Αυτό επιτυγχάνεται με την εφαρμογή της μεθόδου ενεργών ακτίνων (Active Rays), πολικός μετασχηματισμός ενεργών περιγραμμάτων (Active Contours), σε 8 διευθύνσεις και σε δύο κλίμακες του μετασχηματισμού κυματίων (wavelet transform) με φίλτρα Β-spline. Ακολούθως, χρησιμοποιείται μέθοδος επέκτασης περιοχής (region growing) για τον ακριβή προσδιορισμό του περιγράμματος της μικροαποτιτάνωσης. Ως κριτήριο για την αύξηση της περιοχής χρησιμοποιήθηκαν τα σημεία στο περίγραμμα της μικροαποτιτάνωσης, όπως αυτά προσδιορίσθηκαν από τη μέθοδο των ενεργών ακτίνων. Επίσης, υλοποιήθηκε μέθοδος ακτινικής βάθμωσης, η οποία έχει πρόσφατα προταθεί στη βιβλιογραφία για την τμηματοποίηση μικροαποτιτανώσεων, και χρησιμοποιήθηκε για συγκριτική αξιολόγηση. Οι δύο μέθοδοι τμηματοποίησης εφαρμόστηκαν σε 149 ομάδες μικροαποτιτανώσεων, κυρίως πλειόμορφων, που αντλήθηκαν από 130 μαστογραφικές εικόνες από τη βάση DDSM (Digital Database for Screening Mammography). Η ακρίβεια τμηματοποίησης των δύο μεθόδων αξιολογήθηκε από τρεις ακτινολόγους με χρήση 5-βάθμιας κλίμακας. Η ακρίβεια τμηματοποίησης της προτεινόμενης μεθόδου βρέθηκε ίση με 3.96±0.77, 3.97±0.80 και 3.83±0.89, όπως αξιολογήθηκε από κάθε ακτινολόγο, και 2.91±0.86, 2.10±0.94 και 2.56±0.76 για την συγκρινόμενη μέθοδο. Οι διαφορές στην ακρίβεια τμηματοποίησης των δύο μεθόδων ήταν στατιστικώς σημαντικές (Wilcoxon signed-ranks test, p<0.05). Επίσης, μελετήθηκε η επίδραση των δύο μεθόδων τμηματοποίησης στην απόδοση μεθόδου αυτόματης διάγνωσης (χαρακτηρισμό) ομάδων μικροαποτιτανώσεων με υπολογιστή. Η μέθοδος αυτόματης διάγνωσης στηρίζεται σε επιβλεπόμενη ταξινόμηση προτύπων χαρακτηριστικών σχήματος ομάδας αποτιτανώσεων. Συγκεκριμένα, χρησιμοποιήθηκε ταξινομητής ελαχίστων τετραγώνων – ελάχιστης απόστασης και εξήχθησαν χαρακτηριστικά ομοιότητας και διαφοροποίησης (variability) ομάδας μικροαποτιτανώσεων, τα οποία περιγράφουν τη μορφολογία μεμονωμένων αποτιτανώσεων (εμβαδόν, μέγιστη διάμετρος, σχετική αντίθεση). Η απόδοση ταξινόμησης αποτιμήθηκε μέσω εμβαδού καμπύλης παρατηρητών (ROC). Τα χαρακτηριστικά Εμβαδού και μέγιστης Διαμέτρου επέδειξαν σημαντικά υψηλή απόδοση ταξινόμησης (Mann-Whitney U-test, p<0.05) όταν εξήχθησαν από μικροαποτιτανώσεις τμηματοποιημένες με την προτεινόμενη μέθοδο ενεργών ακτίνων (0.82±0.06 και 0.86±0.05, αντίστοιχα). Η απόδοση ταξινόμησης χαρακτηριστικών που εξήχθησαν με μέθοδο τμηματοποίησης ακτινικής βάθμωσης ήταν 0.71±0.08 και 0.75±0.08, αντίστοιχα. Συμπερασματικά, η προτεινόμενη μέθοδος επέδειξε βελτιωμένη ακρίβεια τμηματοποίησης, εκπληρώνοντας ποιοτικά κριτήρια και ενισχύοντας την ικανότητα χαρακτηρισμού των ομάδων αποτιτανώσεων με ανάλυση μορφολογίας (μεγέθους και σχήματος) μεμονωμένων αποτιτανώσεων. Οι περιορισμοί της προτεινόμενης μεθόδου τμηματοποίησης αποδίδονται κυρίως: • Στην ανάλυση δύο κλιμάκων του μετασχηματισμού κυματίου, με αποτέλεσμα τον περιορισμό της προσαρμοστικότητας της μεθόδου σε μικροαποτιτανώσεις διαφορετικών μεγεθών. • Στην μέθοδο επέκτασης περιοχής περιοριζόμενη από σημεία περιγράμματος σε 8 διευθύνσεις. Οι περιορισμοί της αξιολόγησης της προτεινόμενης μεθόδου τμηματοποίησης αποδίδονται κυρίως: • Στην ποιοτική μόνο αξιολόγηση της ακρίβειας τμηματοποίησης, μέσω ανάλυσης παρατηρητών. • Στην χρήση περιορισμένου αριθμού χαρακτηριστικών μορφολογίας στο σύστημα αυτόματης διάγνωσης. Για την αντιμετώπιση των προαναφερθέντων περιορισμών, προτάθηκε η μέθοδος Ενεργών Περιγραμμάτων Πολλαπλών Κλιμάκων με αρχικοποίηση Ενεργών Ακτίνων στην αυτόματα επιλεγόμενη αδρή κλίμακα κυματίου. Αρχικά, χρησιμοποιήθηκε ο μετασχηματισμός συνεχούς κυματίου για την παροχή πολλαπλών κλιμάκων ανάλυσης. Στο πεδίο των πολλαπλών κλιμάκων εντοπίζεται η βέλτιστη αδρή κλίμακα (coarse scale) ανάλυσης με βάση τη μέγιστη απόκριση περιοχής μικροαποτιτάνωσης (scale-space MC signature). Στη συγκεκριμένη βέλτιστη κλίμακα απόκρισης εφαρμόζεται η μέθοδος των ενεργών ακτίνων για τον εντοπισμό σημείων του περιγράμματος της μικροαποτιτάνωσης σε 8 διευθύνσεις. Από αυτά τα σημεία ορίζεται πλήρως το περίγραμμα με χρήση μεθόδου γραμμικής παρεμβολής στη βέλτιστη κλίμακα απόκρισης. Κάθε σημείο του περιγράμματος ακολουθεί την κατεύθυνση μεγιστοποίησης της βάθμωσης εικόνας για τον καθορισμό του περιγράμματος στην βέλτιστη κλίμακα (directional Active Contour). Για την τελική εξαγωγή του περιγράμματος, οι θέσεις των σημείων του περιγράμματος επανακαθορίζονται στις κλίμακες μεγαλύτερης ακρίβειας (fine scales). Η ακρίβεια τμηματοποίησης της δεύτερης προτεινόμενης μεθόδου αξιολογήθηκε ποσοτικά με το κριτήριο επικάλυψης περιοχής. Για το σκοπό αυτό χρησιμοποιούνται τμηματοποιήσεις από ειδικευμένο ακτινολόγο. Τμηματοποιήθηκαν συνολικά 1157 μεμονωμένες μικροαποτιτανώσεις προερχόμενες από 128 ομάδες μικροαποτιτανώσεων, ψηφιοποιημένες σε ανάλυση 50μm (βάση δεδομένων DDSM). Μελετήθηκε επίσης η επίδραση της ακρίβειας τμηματοποίησης της δεύτερης προτεινόμενης μεθόδου στην απόδοση μεθόδου αυτόματης διάγνωσης ομάδων αποτιτανώσεων με βάση χαρακτηριστικά ομοιότητας και διαφοροποίησης (variability) ομάδας μικροαποτιτανώσεων, τα οποία περιγράφουν τη μορφολογία μεμονωμένων αποτιτανώσεων (εμβαδού: εμβαδόν, μέγιστη διάμετρος, σχετική αντίθεση, εκκεντρότητα, συμπαγότητα, διακύμανση ακτινικών αποστάσεων, περιοχής: ροπές 1ης και 2ης τάξης, και περιγράμματος: χαρακτηριστικό ροπής και συχνότητας). Ακολούθως, τέσσερα συστήματα αυτόματης διάγνωσης σχεδιάστηκαν βασιζόμενα στον ταξινομητή ελαχίστων τετραγώνων – ελάχιστης απόστασης και μορφολογικά χαρακτηριστικά εξήχθησαν από τις τρεις αυτόματες μεθόδους τμηματοποίησης (δύο προτεινόμενες και μία συγκρινόμενη). Η ποσοτική αξιολόγηση των προτεινόμενων μεθόδων τμηματοποίησης με χρήση δείκτη επικάλυψης περιοχής απέδειξε ότι μόνο η μέθοδος των Ενεργών Περιγραμμάτων Πολλαπλών Κλιμάκων με αρχικοποίηση Ενεργών Ακτίνων στη βέλτιστη κλίμακα ανάλυσης χαρακτηρίζεται από εξίσου υψηλή απόδοση για τις μικρού (<500μm) και μεγάλου (>500μm) μεγέθους μικροαποτιτανώσεις. Επιπλέον, ο ταξινομητής που βασίστηκε σε χαρακτηριστικά εξαγόμενα από τη βελτιστοποιημένη μέθοδο τμηματοποίησης παρουσίασε καλύτερη απόδοση ταξινόμησης (0.779±0.041) από τους ταξινομητές που βασίστηκαν σε χαρακτηριστικά εξαγόμενα από τη μέθοδο Ενεργών Ακτίνων (0.667±0.041) και τη μέθοδο ακτινικής βάθμωσης (0.670±0.044). Η απόδοση ταξινόμησης του βελτιωμένου αλγόριθμου τμηματοποίησης ήταν δε παρόμοια με την απόδοση του ταξινομητή που βασίστηκε σε χαρακτηριστικά εξαγόμενα από χειροκίνητα τμηματοποιημένες μικροαποτιτανώσεις (0.813±0.037).
Accurate segmentation of microcalcifications (MCs) in mammography is crucial for the quantification of morphologic properties by features incorporated in computer-aided diagnosis (CADx) schemes. At first, a novel segmentation method is proposed implementing active rays (polar-transformed active contours) on B-spline wavelet representation to identify microcalcification contour point estimates in a coarse-to-fine strategy at two levels of analysis. An iterative region growing method is used to delineate the final microcalcification contour curve, with pixel aggregation constrained by the microcalcification contour point estimates. A radial gradient method, representing the current state-of-the-art, was also implemented for comparative purposes. The methods were tested on a dataset consisting of 149 mainly pleomorphic microcalcification clusters originating from 130 mammograms of the DDSM database. Segmentation accuracy of both methods was evaluated by three radiologists, based on a 5-point rating scale. The radiologists’ average accuracy ratings were 3.96±0.77, 3.97±0.80 and 3.83±0.89 for the proposed method, and 2.91±0.86, 2.10±0.94 and 2.56±0.76 for the radial gradient-based method, respectively, while the differences in accuracy ratings between the two segmentation methods were statistically significant (Wilcoxon signed-ranks test, p<0.05). The effect of the two segmentation methods in the classification of benign from malignant microcalcification clusters was also investigated. A Least Square Minimum Distance (LSMD) classifier was employed based on cluster features reflecting three morphological properties of individual microcalcifications (area, length and relative contrast). Classification performance was evaluated by means of the area under ROC curve (Az). The area and length morphologic features demonstrated a statistically significant (Mann-Whitney U-test, p<0.05) higher patient-based classification performance when extracted from microcalcifications segmented by the proposed method (0.82±0.06 and 0.86±0.05, respectively), as compared to segmentation by the radial gradient-based method (0.71±0.08 and 0.75±0.08). The proposed method demonstrates improved segmentation accuracy, fulfilling human visual criteria, and enhances the ability of morphologic features to characterize microcalcification clusters. The limitations of this method could be attributed to: • Multiscale analysis restricted to two scales and ad-hoc selection of the coarsest scale of analysis, limiting the desired size-adaptation property of the proposed segmentation method. • Use of constrained region growing to delineate the final MC region to avoid discontinouities inherent to the 8-contour point estimates. • Segmentation accuracy assessed only qualitatively. • Limited morphology anaysis incorporated into the CADx framework. To overcome these limitations, a second method is introduced adaptive to coarse scale selection to initialize the segmentation algorithm, by means of scale-space signatures. Also, we suggest the analysis in the continuous wavelet transform offering a rich multiscale frame. In this framework, multiscale active contours are introduced, utilizing as initial step the previously proposed Active Rays method combined to linear interpolation, for initial contour estimation. Then, each contour point follows the direction where the gradient is maximized. MCs are finally delineated by refining each contour point position at finer scales more accurately. Segmentation accuracy is quantitatively assessed by means of the Area Overlap Measure, utilizing manual segmentation of individual MCs as ground truth, provided by expert radiologists. A total of 1157 individual MCs were segmented in a dataset of 128 MC clusters, digitized at 50μm pixel resolution. To further ensure feature reliability, features extracted from the improved segmentation method were compared (Pearson correlation) to features extracted from manual experts’ delineations. Following, four CADx schemes were implemented utilizing Least Square Minimum Distance (LSMD) classifier and morphology features extracted from the two proposed and the Radial Gradient method. Training of all classifiers was accoblished by features extracted from manually segmented MCs. Quantitative analysis indicated that Multiscale Active Contour method initialized by Active Rays (MAC-AR) had similar Area Overlap Measure performance either for small and large MCs. Furthermore, the improved method demonstrated better performance in terms of classification performance (Az=0.78), as compared to Multiscale Active Rays constrained Region Growing (MAR-RG) (Az=0.67) and the radial gradient one (Az=0.67), however, statistically similar to manual segmentation, representing the best performance (Az=0.81).

碩士
中華大學
電機工程學研究所
86
Proximal dental caries detection is a clinical challenge. Among various caries examinations, proxial caries are easily missed in physical examinations, whichare the most frequently used method. Radiological examination is considerde by many as th most accurate and practical proximal caries examination method. However, sensitivity and specificity of proximal caries detection on radiographs have been demonstrated to vay srignificantlh among different observers. Experienced oral radiologists and dentists can accurately detect more proximal caries, but these trained personnel are few in the profession. There is a need for a more objective proximal caries detection method, and Computer-Assisted Diagnosis System (CADS ) has been reported as a promising one. 　　We proposed a two stage CADS for proximal caries detection.The first stage is to locate an individual tooth, and the second stage is to detect the possible caries regions on this individual tooth. In the first stage, we manually select individual tooth, and used auto-thresholding technique to remove background. Then we used curve fitting to generate virtual proximal edge, at the same time removed the non-target teeth. In the second stage, we combine the local mean information and the unit normal vectors of tooth surface to locate candidate proxinal carious regions. Afterward, the horizontally cumulative sum of each point on the proximal surve is computed to find possible caries regions. The developed CADS could detect proximal caries on eight typical images from The Dentistry Department of National Taiwan University Hospital.

碩士
國立臺灣大學
資訊網路與多媒體研究所
103
Liver cancer is the tenth most common cancer in the recent years in USA. Therefore, early detection and well treatment are very important for the patient. Computed tomography (CT) is one of the most common and robust imaging techniques for the detection of liver. CT scanners allow multiple-phase sequential scans of the whole liver to be obtained during the injection of contrast material. In this paper, the main purpose is to build a computer-aided diagnosis (CAD) system to extracted features from tumors and diagnose the liver tumor in multiple-phase CT. There are two kinds of data in this paper, one is the four phase CT images and the other is three phase CT images. The experiment of two kinds of data will do in the same way but separately. In the proposed CT computer-aided diagnosis (CAD) system, the tumor was indicated by user and the tumor was segmented by a region growing algorithm. After tumor segmentation, three kinds of features were extracted from the tumor including texture features, shape features, and kinetic curve features. The texture features quantify 3 dimensions (3-D) texture information of tumor based on the grey level co-occurrence matrix. Compactness, margin and elliptic model were used to describe the 3-D shape information of tumor in the shape features. The last kind of features is the kinetic curve features which was extracted from each phase of tumor and represent the intensity variation between each phase. By analyzing the three kinds of features in the three phase and four phase CT images, we have two experiment results. In the experiment of four phase CT images, 40 tumors with 29 benign and 11 malignant tumors were used in this CAD system to evaluate the performance. The accuracy, sensitivity, specificity, and AZ were up to 77.5% (31/40), 72.73% (8/11), 79.31% (23/29), and 0.7791, respectively. In the experiment of three phase CT images, 71 tumors with 49 benign and 22 malignant tumors were used in this CAD system to evaluate the performance. The accuracy, sensitivity, specificity, and AZ are up to 81.69% (58/71), 81.82% (18/22), 81.63% (40/49), and 0.8713. As a result, the accuracy, sensitivity, specificity, and AZ are better in the experiment of three phase CT images than four phase CT images.