| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 4.44 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A ressonância magnética (RM) representa um dos mais significativos avanços tecnológicos no que se refere a diagnósticos médicos por imagem. Dependendo do exame que se pretende realizar, pode ser necessário o uso do gadolínio (como agente de contraste exógeno) para se obter imagens com maior valor de diagnóstico. O contraste de gadolínio realça a diferença de sinal entre tecidos saudável e patológico, o que permite o diagnóstico mais preciso e tomada de decisões do médico. Porém, o uso do contraste intravenoso pode ser prejudicial a pessoas com, por exemplo, insuficiência renal, não sendo, neste caso, recomendado o seu uso. Pessoas com este tipo de deficiência podem desenvolver fibrose sistémica nefrogénica (FSN), que não tem cura e pode ser letal. Neste contexto, este trabalho teve como objectivo verificar se há uma correlação entre as imagens ponderadas em T1 e T2 para simular, a partir de suas informações, imagens artificialmente contrastadas. Uma vez que as relações entre os valores dos pixels das diferentes regiões e tecidos observados em imagens de RM são complexas, esta análise utilizou como ferramenta redes neuronais artificiais (RNA), um sistema computacional com capacidade de aprender relações complexas dentro de um conjunto de dados, processar e desempenhar uma determinada tarefa. Os testes foram realizados com imagens segmentadas em 7, 14 e 32 níveis de cinzento. As imagens artificialmente contrastadas, geradas pela RNA, foram comparadas às imagens reais de pós contraste (segmentada no mesmo número de cinzentos a testar) de cada paciente, no intuito de verificar não só a capacidade da RNA em gerar as imagens pretendidas, mas também a influência que o número de cinzentos da imagem exerce sobre o resultado final. Como resultado, a rede neuronal obteve uma média de 54,22% de desempenho na aprendizagem e de 66,05% na eficácia em gerar, como resposta ao treino, uma imagem semelhante à pretendida (média de semelhança obtida entre as imagens comparadas), o que permite concluir que esta correlação não apenas pode ser avaliada como também artificialmente simulada.
ABSTRACT - Magnetic resonance imaging (MRI) represents one of the most significant technological advances in medical imaging diagnostics. Depending on the intended examination, the use of gadolinium (as an exogenous contrast agent) may be necessary to obtain images with a higher diagnostic value. The gadolinium contrast enhances the difference in signal between healthy and pathological tissues, which allows more accurate diagnosis and decision making of the doctor. However, the use of intravenous contrast may be harmful to people with, for example, renal failure, and in this case its use is not recommended. People with this type of disability may develop a condition called nephrogenic systemic fibrosis (NSF), which has no cure and can be fatal. In this context, this study aimed to verify if there is a correlation between T1 - weighted images and T2 - weighted images to simulate, from their information, artificially contrasted images. Since the relationships between the pixel values of the different regions and tissues observed in MRI images are complex, this analysis used as an Artificial Neural Networks (ANN) tool, a computational system capable of learning complex relationships within a set of data, process and perform a certain task. The tests were performed with segmented images at 7, 14 and 32 levels of gray. The artificially contrasted images, generated by the ANN, were compared to the real post-contrast images (segmented in the same number of gray to be tested) of each patient, in order to verify not only the ANN's ability to generate the desired images, but also the influence that the number of grays of an image exerts on the final result. As a result, the neural network obtained an average of 54.22% of performance in learning and 66.05% of effectiveness in generating, in response to training, an image similar to that intended (average similarity obtained between the compared images), the which allows us to conclude that this correlation can not only be evaluated but also artificially simulated.
ABSTRACT - Magnetic resonance imaging (MRI) represents one of the most significant technological advances in medical imaging diagnostics. Depending on the intended examination, the use of gadolinium (as an exogenous contrast agent) may be necessary to obtain images with a higher diagnostic value. The gadolinium contrast enhances the difference in signal between healthy and pathological tissues, which allows more accurate diagnosis and decision making of the doctor. However, the use of intravenous contrast may be harmful to people with, for example, renal failure, and in this case its use is not recommended. People with this type of disability may develop a condition called nephrogenic systemic fibrosis (NSF), which has no cure and can be fatal. In this context, this study aimed to verify if there is a correlation between T1 - weighted images and T2 - weighted images to simulate, from their information, artificially contrasted images. Since the relationships between the pixel values of the different regions and tissues observed in MRI images are complex, this analysis used as an Artificial Neural Networks (ANN) tool, a computational system capable of learning complex relationships within a set of data, process and perform a certain task. The tests were performed with segmented images at 7, 14 and 32 levels of gray. The artificially contrasted images, generated by the ANN, were compared to the real post-contrast images (segmented in the same number of gray to be tested) of each patient, in order to verify not only the ANN's ability to generate the desired images, but also the influence that the number of grays of an image exerts on the final result. As a result, the neural network obtained an average of 54.22% of performance in learning and 66.05% of effectiveness in generating, in response to training, an image similar to that intended (average similarity obtained between the compared images), the which allows us to conclude that this correlation can not only be evaluated but also artificially simulated.
Description
Mestrado em Radiações Aplicadas às Tecnologias da Saúde. Ramo de especialização: Ressonância Magnética.
Keywords
Ressonância magnética Contraste de Gadolínio Rede neuronal artificial Imagem de contraste Magnetic resonance Contrast of Gadolinium Artificial neural network Contrast image
Citation
Nunes RO. Desenvolvimento de um sistema de reconhecimento de padrões para simular imagens de ressonância magnética com contraste [dissertation]. Lisboa: Escola Superior de Tecnologia da Saúde de Lisboa/Instituto Politécnico de Lisboa; 2018.
Publisher
Escola Superior de Tecnologia da Saúde de Lisboa/Instituto Politécnico de Lisboa