A Ressonância Magnética (MRI - do inglês Magnetic Ressonance Imaging) é uma modalidade de imagem médica poderosa, difundida e indispensável. O ACR (American College of Radiology) recomenda que o desempenho das máquinas de ressonância magnética seja monitorado, repetindo os testes de qualidade de imagem a cada 7 dias ou menos. Testes de qualidade são realizados em imagens de um objeto de geometria e composição conhecidas, denominado phantom. Uma máquina em bom estado deve gerar imagem que retrate a anatomia sob inspeção com as dimensões e características corretas, e permitir a detecção de pequenos furos em condições de baixo contraste. Alguns métodos automatizados foram propostos na literatura, mas a automação de dois dos testes do ACR, de alto e baixo contraste, continua sendo um problema em aberto. Esta tese apresenta uma proposta para automatizar o teste de baixo contraste do ACR. Este teste, geralmente, é feito por técnicos analisando a imagem de phantom. No entanto, a análise automatizada seria capaz de reduzir custos, melhorar a repetibilidade e confiabilidade das medidas de controle de qualidade. Os trabalhos sobre automação dos testes de baixo contraste do ACR são escassos e, até onde sabemos, nenhum deles produziu resultados robustos o suficiente que permita substituir o trabalho humano. Podemos separar esta tese em duas fases principais. Na primeira, consideramos as respostas dos técnicos seniores, com mais de 10 anos de experiência, como nosso padrão ouro. Utilizamos um banco de dados com 620 conjuntos de imagens phantom ACR, que foram adquiridos em máquinas de diferentes fornecedores, campos e bobinas, totalizando 74.400 furos de baixo contraste. Técnicos com mais de 10 anos de experiência rotularam cada furo como visível ou invisível. Os algoritmos de aprendizado de máquina foram alimentados com características obtidas manualmente, com objetivo de extrair informações dos furos e seus arredores. Entre os cinco métodos testados, a regressão logística apresentou a maior área sob a curva ROC (0,878) e o maior alfa de Krippendorff (0,995). Os resultados alcançados nesta fase do trabalho já são substancialmente melhores do que os relatados anteriormente na literatura. Também são melhores do que as classificações feitas por técnicos juniores, com menos de 5 anos de experiência. Estes primeiros resultados já são um indicativo de que o teste de resolução de baixo contraste, ACR MRI, pode ser automatizado usando as técnicas de aprendizado de máquina.

Magnetic Resonance Imaging (MRI) is a powerful, widespread and indispensable medical imaging modality. The American College of Radiology (ACR) recommends that the performance of MRI machines be monitored by repeating the image quality tests every 7 days or less. Quality tests are performed on images of an object of known geometry and composition called phantom. A machine in good condition must generate an image that depicts the anatomy under inspection with the correct dimensions and characteristics; and allow the detection of small structures under low contrast conditions. Some automated methods have been proposed in the literature, but the automation of two of the ACR tests, high and low contrast, remains an open problem. This thesis presents a proposal to automate the ACR low contrast test. This test is usually done by technicians analyzing the phantom image, but automated analysis would reduce costs, improve repeatability and reliability of quality control measures. Work on automating ACRs low-contrast tests is scarce and, as far as we know, none of them has produced results robust enough to replace human work. We can separate this project into two main phases. In the first one, we consider the answers of senior technicians, with more than 10 years of experience, as our gold standard. We used a database with 620 sets of ACR phantom images that were acquired on machines from different vendors, fields and coils, totaling 74,400 low-contrast structures. Technicians with more than 10 years of experience labeled each structure as detectable or undetectable. Machine learning algorithms were fed manually designed features to extract information from structures and their surroundings. Among the five methods tested, Logistic Regression presented the largest area under the ROC curve (0.878) and the largest Krippendorff alpha (0.995). The results achieved in this phase of the work are already substantially better than those previously reported in the literature. They are also better than classifications made by junior technicians (less than 5 years of experience). These early results are already an indication that the ACR MRI low-contrast resolution test can be automated using machine learning techniques. Deep learning is part of a family of machine learning methods based on artificial neural networks. Among deep learning techniques, convolutional neural networks have been used in image classification, producing results comparable to, and in some cases superior to, those of human experts. In the second phase, we use convolutional neural networks to emulate the detection of low contrast structures (holes) in an ACR phantom. Carefully analyzing our dataset, we concluded that senior technicians make as many mistakes as less experienced technicians, and therefore years of experience do not, by themselves, guarantee greater accuracy in the classification task. Thus, in the second phase, we changed the gold standard from hole visibility to the median of the responses of all technicians (regardless of years of experience) who did not make gross mistakes in classifying the holes in the image. We used a subset of 100 phantom ACR acquisitions from the previous phase, totaling 12,000 holes. For statistical robustness, we repeated training and testing 5 times, using 5- fold cross validation. We obtained an average AUC (Area Under the ROC Curve) of 0.983±0.003 and an average accuracy of 93.2±0.7% at the EER point (Equal Error Rate). Applying the model obtained to a completely independent test dataset with 10,800 structures, we obtained an AUC of 0.979. The predictions of our model in the classification of spokes (sets of 3 holes) agree in 93.83% of cases with the median of the technicians answers. These results are better than the answers of any individual technician. We conclude that our system can replace the human technician in the ACR low-contrast test and can still provide real-time answers to help in the training of new technicians involved in the process.