Algoritimo de detecção de retinopatia diabética baseado em aprendizado de máquina

• Autor(a) principal: Reis, Leonardo Patrocínio dos
• Data de Publicação: 2023
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/20.500.14289/17725
• topic: Retinopatia diabética; Classificação de imagens médicas; Aprendizado de máquina; Redes neurais; U-Net; Mecanismo de atenção; Imagens de fundo de olho; Detecção precoce; Segmentação de veias; Diabetic retinopathy; Medical image classification; Machine learning; Neural network; Inception; Mechanism of Attention; Fundus imaging; Early detection; Vein segmentation; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO; ENGENHARIAS::ENGENHARIA ELETRICA::CIRCUITOS ELETRICOS, MAGNETICOS E ELETRONICOS

In this study, the application of neural network and machine learning techniques was explored in order to identify the presence of lesions related to diabetic retinopathy (DR) in fundus images. DR is a frequent complication in diabetic individuals and can lead to vision loss if not detected and treated in a timely manner. The architecture of the classification model proposed in this work is composed of two decision streams that are concatenated to generate the final classification. The first flow uses a U-Net network to segment and extract veins and blood vessels from the original image, followed by an Inception model with an attention mechanism for classification. The second stream directly processes the raw image through an Inception model with an attention mechanism. The proposed model was trained and validated using three combined public datasets (ARIA, RFMiD and STARE). Tools employed in development included Python, TensorFlow, Keras, OpenCV and other complementary libraries. The final model reached an accuracy of 95.4% and a sensitivity of 94.87% in classifying diabetic retinopathy lesions, demonstrating its potential to contribute to the early detection and adequate treatment of this ocular complication.