Classifica??o de amiloidose em imagens digitais de bi?psias renais utilizando corantes n?o espec?ficos

• Ano de defesa: 2023
• Autor(a) principal: Oliveira, Gledson de
• Orientador(a): Duarte, Angelo Am?ncio
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Feira de Santana
• Programa de Pós-Graduação: Programa de P?s-Gradua??o em Ci?ncia da Computa??o
• Departamento: DEPARTAMENTO DE TECNOLOGIA
• País: Brasil
• Palavras-chave em Português: Patologia Computacional; Amiloidose; Desbalanceamento de Classe
• Palavras-chave em Inglês: Computational Pathology; Amyloidosis; Class Imbalance
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA
• Link de acesso: http://tede2.uefs.br:8080/handle/tede/1743

Computational Pathology is a field of study that utilizes computational methods to assist medical pathologists in the analysis of pathological images, employing machine learning algorithms to contribute to faster and more precise diagnoses. Nevertheless, there is still much to be researched, as exemplified by the case of amyloidosis, a rare condition that poses a challenge to the development of efficient classifiers for the condition due to a limited amount of available images for training automatic classifiers. Additionally, amyloidosis presents an additional complication arising from the necessity of using specific dyes for lesion detection by physicians, further reducing the pool of available images. Based on this research problem, this study employed an approach utilizing classical convolutional neural network models to construct an automatic amyloidosis classifier, trained using colored images with non-specific dyes for the lesion. Initially, these models were trained using an imbalanced dataset with fewer images for amyloidosis to establish a research baseline. Subsequently, data balancing techniques such as Random Undersampling and Random Oversampling, along with Ensemble-Based algorithms, were applied to address class imbalance. As a result of this work, models capable of identifying the lesion with a false negative rate of up to 4.5% were obtained, with better performance observed for the Inception model when trained with the RUS dataset and for the Ensemble-RUS model.