Deep learning and data warehousing techniques applied to real data in the medical domain

• Ano de defesa: 2023
• Autor(a) principal: Lima, Daniel Mário de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizagem profunda; Armazém de dados; Cardiac MRI; Clinical research; Data warehouse; Deep learning; Dermatoscopia; Dermatoscopy; Pesquisa clínica; RM cardíaca
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-01092023-164636/

This study aims to increase the use of medical data and the ability to automated diagnosis through the integration and homogenization of the databases from the SI3 Health Information System of the Heart Institute (InCor / HC.FMUSP), and investigate the application of state-ofthe- art machine learning models known as Deep Learning, assessing the potential of Deep Learning to computerized diagnosis. As results, a database was prepared for clinical research in the OMOP-CDM format, called InCor-CDM. In the second study we obtained up to 91% overall accuracy in the classification of cutaneous lesions using a deep convolutional neural network on the ISIC database of dermatoscopic images. In the third paper we improved the segmentation of heart magnetic resonance images, on average, by 1.7% in the Dice metric and 2.5x in the training speed of a U-Net convolutional neural network using a localization algorithm. These results demonstrate steps of data preparation; deep learning applied to high-level medical concepts multi-classification for diagnosis; and deep learning applied to low-level image data Cardiac MRI image segmentation.