| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.44 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Em 2020, estima-se que 325.000 pessoas tenham sido diagnosticadas com melanoma, estimando-se que 50.000 mortes em todo o mundo por doença. Além disso, o melanoma é um dos cancros mais comuns em jovens adultos. O melanoma, quando em fase inicial, tem uma grande taxa de sobrevivência, com cerca de 97%, no entanto, em fases tardias a taxa de sobrevivência é apenas de cerca de 10%, o que implica que uma deteção precoce é extremamente importante. Nos últimos anos, o machine learning tem feito progressos notáveis e com isso seguiu-se o número de pesquisas e utilizações em questões médicas. Seria possível utilizar os grandes avanços na inteligência artificial e na classificação de imagem para criar uma aplicação para detetar precocemente o melanoma? Nesta tese são testadas três arquiteturas de machine learning para prever melanoma, VGG16, VGG19 e MobileNet V2. Todas as três arquiteturas usadas foram pré-treinadas e aplicadas a um dataset que compila todos os principais datasets de melanoma públicos. A arquitetura com melhores resultados foi o MobileNet V2, que com afinação alcançou mais de 96% de precisão. Foi também testado para aplicar o préprocessamento nas imagens do dataset (filtro CLAHE), que acabou por ter uma precisão ligeiramente menor na mesma arquitetura, com menos 0,2%. Com o melhor modelo, foi criada uma aplicação simples para permitir aos pacientes utilizarem o modelo de machine learning. O modelo de aprendizagem automática criado nesta tese alcançou uma precisão excecionalmente elevada, em comparação com a literatura. Além disso, destaca-se por formar o resto porque utiliza uma arquitetura recente e leve e tem uma aplicação em execução na parte frontal que permite a todos, mesmo pessoas sem conhecimento técnico, utilizar a aplicação.
In 2020, there were an estimated 325,000 people diagnosed with melanoma, and an estimated 50,000 worldwide deaths from the disease. Moreover, melanoma it’s one of the most common cancers in young adults. Melanoma, when in early stages, has a great survival rate, with around 97%, however, in late stages the survival rate is only around 10%, which implies that an early detection is extremely important. In recent years, deep learning has made outstanding breakthroughs and with that the number of researches and uses in medical issues followed. Would it be possible to use the great advances in artificial intelligence and image classification to create an application to early detect melanoma? In this thesis three machine learning architectures are tested to predict melanoma, VGG16, VGG19 and MobileNet V2. All three architectures used were pre-trained and applied to one dataset that compiles all major public melanoma datasets. The architecture with best results was the MobileNet V2, that with fine-tuning achieved more than 96% of accuracy. It was also tested to apply pre-processing on the images of the dataset (CLAHE filter), which turned out to have a slightly lower accuracy on the same architecture, with less 0,2%. With the best model, a simple application was created to allow the patients to use the machine learning model. The machine learning model created in this thesis achieved an exceptionally high accuracy, compared with the literature. Also, it stands out form the rest because it uses a recent and lightweight architecture and has an application running on the frontend which allows everybody, even people without technical knowledge, to use the application.
In 2020, there were an estimated 325,000 people diagnosed with melanoma, and an estimated 50,000 worldwide deaths from the disease. Moreover, melanoma it’s one of the most common cancers in young adults. Melanoma, when in early stages, has a great survival rate, with around 97%, however, in late stages the survival rate is only around 10%, which implies that an early detection is extremely important. In recent years, deep learning has made outstanding breakthroughs and with that the number of researches and uses in medical issues followed. Would it be possible to use the great advances in artificial intelligence and image classification to create an application to early detect melanoma? In this thesis three machine learning architectures are tested to predict melanoma, VGG16, VGG19 and MobileNet V2. All three architectures used were pre-trained and applied to one dataset that compiles all major public melanoma datasets. The architecture with best results was the MobileNet V2, that with fine-tuning achieved more than 96% of accuracy. It was also tested to apply pre-processing on the images of the dataset (CLAHE filter), which turned out to have a slightly lower accuracy on the same architecture, with less 0,2%. With the best model, a simple application was created to allow the patients to use the machine learning model. The machine learning model created in this thesis achieved an exceptionally high accuracy, compared with the literature. Also, it stands out form the rest because it uses a recent and lightweight architecture and has an application running on the frontend which allows everybody, even people without technical knowledge, to use the application.
Description
Keywords
Melanoma Inteligência artificial Machine learning Classificação de imagens Segmentação de imagens Supervised learning Imagem médica MobileNet V2 Python Artificial Intelligence Machine learning Image classification Image segmentation Supervised learning Medical imaging