Combining heterogeneous data and deep learning models for skin cancer detection

• Ano de defesa: 2020
• Autor(a) principal: Pacheco, André Georghton Cardoso
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Ciência da Computação; Centro Tecnológico; UFES; Programa de Pós-Graduação em Informática
• Palavras-chave em Português: Deep Learning; Data Aggregation; Ensemble of Deep Models; Convolutional Neural Networks; Image Classification; Skin Cancer Detection; subject.br-rjbn; Ciência da Computação
• Link de acesso: http://repositorio.ufes.br/handle/10/14433

Currently, Deep Neural Networks (DNN) are the most successful and common methodologies to tackle medical image analysis. Despite the success, applying Deep Learning for these types of problems involves several challenges such as the lack of large training datasets, data variance, and noise sensitivity. In this thesis, our main focus is on proposing solutions to assist Deep Learning models to deal with these issues when they are applied to medical (clinical) image problems, in particular for skin cancer detection. Basically, we work on two main topics: data classification using images and context meta-data and dynamic weighting for an ensemble of deep models. First, we propose two methods to combine images and meta-data; one method is based on features concatenation that uses a mechanism to balance the contribution of each source of data; the second method, named Meta-data Processing Block (MetaBlock), uses meta-data to support the classification by identifying the most relevant features extracted from the images. Next, we propose an approach, based on a Dirichlet distribution and Mahalanobis distance, to learn dynamic weights for an ensemble of deep models. The learned weights are used to reduce the impact of weak models on the aggregation operator and to online select models from the ensemble. All these methods are evaluated in well-known image classification datasets in different experiments. Results show that the proposed methods are competitive with other approaches that deal with the same problems. Lastly, we carry out a case study using a new skin lesion dataset – composed of clinical images collected from smartphones and patient demographics – collected in partnership with the Dermatological and Surgical Assistance Program of the Federal University of Espírito Santo. Results achieved using this dataset are comparable to other recent performance reported in the literature, which shows that the proposed algorithms are viable to deal with skin cancer detection.