Deep learning for melanoma classification : a study using skin lesion images

• Autor(a) principal: Silva, Miguel Marinho Ferreira da
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/42247
• topic: Cutaneous melanoma; Deep learning; Convolutional neural networks; Fβ score; Medical imaging; Melanoma cutâneo; Aprendizagem profunda; Redes neuronais convolucionais; Pontuação Fβ; Imagiologia médica; Domínio/Área Científica::Ciências Sociais::Economia e Gestão

Cutaneous melanoma is considered the skin cancer with highest mortality rate and has been gaining the attention of the medical community due to its rapidly increasing incidence. Advancements in computational technologies have paved the way for innovative image detection methods that can be transferable to medical ap plications, significantly enhancing the potential for early intervention in melanoma diagnosis. To make diagnosis more accurate and to further increase survival rates, this study employs deep learning techniques on an extensive dataset derived from multiple sources. Utilizing Microsoft Azure Cloud as the computational infras tructure, trial and error approach was employed by hyperparameterizing several convolutional neural networks (CNN) where the decision criteria were choosing the one with highest Fβ Score. MAR-MELA-CNN is an innovative ensemble model in corporating six state-of-the-art pre-trained CNN architectures: Xception, VGG16, ResNet50, NASNetMobile, MobileNetV2, and InceptionV3. The primary goal of this research is to further understand CNN’s efficiency in the diagnosis of melanoma and to furthermore measure its performance on a merged dataset. The proposed algorithm achieved a Fβ score of 85%, an area under the curve (AUC) score of 93%, and an average precision (AP) score of 92%, promising diagnostic tool for cutaneous melanoma compared to traditional methods. Further improvements lay in the improvement of the architecture, expansion of the computational instances as well as of the dataset. Another field of future work could be devising a strategy for real-time implementation of this model in a hospital setting, as it could be of vital importance to provide swift support to doctors.