With the rapid development of computer and information technologies, medical imaging has become one of the major sources of information for therapy and research in medicine, biology and other fields. Along with the advancement of medical imaging techniques, computer-aided detection and diagnosis (CAD/CADx) has recently emerged to become one of the major research subjects within the area of diagnostic radiology and medical image analysis. This thesis presents two multi-level learning-based approaches for medical image understanding with applications of CAD/CADx. The so-called "multi-level learning strategy" relies on that supervised and unsupervised statistical learning techniques are utilized to hierarchically model and analyze the medical image content in a "bottom up" way. As the first approach, a learning-based algorithm for automatic medical image classification based on sparse aggregation of learned local appearance cues is proposed. The algorithm starts with a number of landmark detectors to collect local appearance cues throughout the image, which are subsequently verified by a group of learned sparse spatial configuration models. In most cases, a decision could already be made at this stage by simply aggregating the verified detections. For the remaining cases, an additional global appearance filtering step is employed to provide complementary information to make the final decision. This approach is evaluated on a large-scale chest radiograph view identification task and a multi-class radiograph annotation task, demonstrating its improved performance in comparison with other state-of-the-art algorithms. It also achieves high accuracy and robustness against images with severe diseases, imaging artifacts, occlusion, or missing data. As the second approach, a learning-based approach for automatic segmentation of ill-defined and spiculated mammographic masses is presented. The algorithm starts with statistical modeling of exemplar-based image patches. Then, the segmentation problem is regarded as a pixel-wise labeling problem on the produced mass class-conditional probability image, where mass candidates and clutters are extracted. A multi-scale steerable ridge detection algorithm is further employed to detect spiculations. Finally, a graph-cuts technique is employed to unify all outputs from previous steps to generate the final segmentation mask. The proposed method specifically tackles the challenge of inclusion of mass margin and associated extension for segmentation, which is considered to be a very difficult task for many conventional methods. / Master of Science

Made available in DSpace on 2015-05-14T12:36:35Z (GMT). No. of bitstreams: 1 arquivototal.pdf: 4611703 bytes, checksum: c5e21f035f26ff6959248fbd243a0217 (MD5) Previous issue date: 2012-04-20 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES / Breast cancer is the most common cause of death by cancer in the female population and a serious world health problem. The mammographic exam allows an early detection which reduces the mortality rate of the disease. Its efficiency has made it the standard procedure for breast cancer diagnosis. These reasons have led to the development of Computer-Aided Detection and Diagnosis (CADDx) systems that assist the physician by working as a second opinion in the diagnostic. One of the algorithms studied during the development of this work, the mass detection algorithm created by Ozekes et al, has shown great potential reaching 99% of sensibility when applied in the test group images. However, its many parameters and the need to manual calibrate them make it impossible to use it in the constructions of practical CADDx systems. This work presents an automatic method for mass detection in mammography based on the algorithm of Ozekes et al. Multilayer Perceptron artificial neural networks (ANN) are used as functional approximators to automatically calibrate the necessary parameters of the proposed method. The computation of the neural networks produces the values used as parameters for thresholding and template application stages. Feature selection and network topologies were chosen by means of empirical tests. Results show in its best configuration point 82% of sensibility and 7,51 false positives per image. After a false positive reduction, 74% of sensibility and 3,56 false positives per image were achieved. Future works include the study of a wider set of image features and preprocessing algorithms. / O câncer de mama é a causa mais comum de morte por câncer na população feminina e um sério problema de saúde mundial. A mamografia permite uma detecção precoce do câncer, reduzindo a mortalidade da doença. Sua eficiência tornou-a procedimento padrão para diagnóstico do câncer de mama. Essas razões levaram ao desenvolvimento de sistemas computadorizados para o auxílio à detecção e ao diagnóstico - em inglês, Computer-Aided Detection and Diagnosis (CADDx) - que auxiliam os profissionais da saúde provendo uma segunda opinião ao diagnóstico. Um dos algoritmos estudados durante o desenvolvimento do trabalho, o algoritmo para detecção de massas criado por Ozekes et al, mostrou grande potencial atingindo 99% de sensibilidade quando aplicado nas imagens testadas. Entretanto, seus muitos parâmetros, e a calibração manual de cada um deles, tornam impossível a aplicação do algoritmo na construção de sistemas CADDx reais. Esse trabalho apresenta um método automático para detecção de massas em mamografias baseado no algoritmo de Ozekes et al. Redes neurais artificiais (RNA) Perceptron multicamadas são usadas como aproximadores universais para a calibração dos parâmetros necessários ao método. A computação dessas redes produz os valores que deverão ser usados como parâmetros para as etapas de binarização e aplicação dos templates. A seleção de atributos e topologias das redes neurais foi definida empiricamente. Resultados mostram, na melhor configuração do sistema, 82% de sensibilidade 7,51 falsos positivos por imagem e, após uma redução de falsos positivos, 74% de sensibilidade e 3,56 de falsos positivos por imagem. Trabalhos futuros incluem o estudo de mais atributos e descritores de imagens além da experimentação de outros algoritmos para pré-processamento.