Detec??o autom?tica de glom?rulos em imagens histol?gicas renais digitais

• Ano de defesa: 2019
• Autor(a) principal: Rehem, Jonathan Moreira Cardozo
• Orientador(a): Angelo, Michele F?lvia
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Feira de Santana
• Programa de Pós-Graduação: Mestrado em Computa??o Aplicada
• Departamento: DEPARTAMENTO DE CI?NCIAS EXATAS
• País: Brasil
• Palavras-chave em Português: Deep learning; Redes Neurais Convolucionais; Glom?rulos; Detec??o de objetos
• Palavras-chave em Inglês: Deep learning; Convolutional Neural Networks; Glomeruli; Object Detection
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://tede2.uefs.br:8080/handle/tede/1342

Glomerulopathies, kidney diseases, affect thousands of people in Brazil and in the entire world, this number is growing constantly. The glomeruli are microscopic structures present in kidney and your examination by a doctor determines the kind and the degree of kidney disease. Kidney tissue images can be scanned or photographed, enabling the computational processing. Nowadays, detection and segmentation are made manually by a pathologist doctor. Thus, this research aims at propose a glomeruli automatic detection method on histological digital kidney tissue images. For this, we use deep learning techniques to train capable models to automate this task. Digital images photographed in varied approximation scales was used to compose train and test datasets. Tensorflow Object Detection API (Application Programming Interface) framework was used implements, train and test the models SSD (Single Shot Detection) Inception V2(SI2) and Faster RCNN (Region-based Convolutional Neural Network) Inception V2 (FRI2). Reaching 0.8831 mAP - 0.94 F1 Score when using the SI2 model, 0.8723 mAP and 0.97 F1 Score when utilizing FRI2 model. The SI2 model. The SI2 model is the most efficient for this task because it is 64% faster in training time and 98% faster in detecting glomeruli in each image. This work demonstrate the efficiency of deep learning techniques as solution for this problem, advancing the improvement of techniques for gloeruli automated detection.