Publicação

Deep learning applied to the classification of skin lesions

Ver documento

Detalhes bibliográficos
Resumo:Skin cancer has been a global health issue and its diagnosis is a challenge in the medical field. Among all the types of skin cancer, melanoma is the worst and can be lethal if not early treated. The use of deep learning techniques, specifically, convolutional neural networks can help to improve the accuracy and speed up the classification of skin lesions. In this work, we aim to employ different image preprocessing techniques, various convolutional neural network models, data augmentation, and ensemble techniques to compare their results and provide an analysis of the data obtained. To achieve that, it was performed several experiments combining different image preprocessing techniques, which, paired with data augmentation strategies, aim to enhance the accuracy and reliability of the classification models. Additionally, three ensemble methods were tested to improve the classification systems’ robustness and reliability by gathering the strengths of each model. Our best result was the ensemble of EfficientNet-B2, EfficientNet-B5, and ResNeSt101 models with the application of data augmentation, and the combination of color constancy and hair removal techniques. This combined approach achieved a balanced accuracy of0.8132. By offering insights into the challenges faced, methodologies employed, and results obtained, this story aims to serve as a guide for researchers and practitioners aiming to advance the field of skin lesion classification using deep learning. Keywords: Deep Learning; Skin Lesion Classification; Image preprocessing.
Autores principais:Silva, Giuliana Martins
Assunto:Deep learning Skin lesion classification Image preprocessing
Ano:2023
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Instituto Politécnico de Bragança
Idioma:inglês
Origem:Biblioteca Digital do IPB
Descrição
Resumo:Skin cancer has been a global health issue and its diagnosis is a challenge in the medical field. Among all the types of skin cancer, melanoma is the worst and can be lethal if not early treated. The use of deep learning techniques, specifically, convolutional neural networks can help to improve the accuracy and speed up the classification of skin lesions. In this work, we aim to employ different image preprocessing techniques, various convolutional neural network models, data augmentation, and ensemble techniques to compare their results and provide an analysis of the data obtained. To achieve that, it was performed several experiments combining different image preprocessing techniques, which, paired with data augmentation strategies, aim to enhance the accuracy and reliability of the classification models. Additionally, three ensemble methods were tested to improve the classification systems’ robustness and reliability by gathering the strengths of each model. Our best result was the ensemble of EfficientNet-B2, EfficientNet-B5, and ResNeSt101 models with the application of data augmentation, and the combination of color constancy and hair removal techniques. This combined approach achieved a balanced accuracy of0.8132. By offering insights into the challenges faced, methodologies employed, and results obtained, this story aims to serve as a guide for researchers and practitioners aiming to advance the field of skin lesion classification using deep learning. Keywords: Deep Learning; Skin Lesion Classification; Image preprocessing.