Sistema de apoio à decisão baseado em Deep Learning para dianóstico de pneumonia em radiografias torácicas

Abstract

Pneumonia is one of the leading causes of global mortality, especially among children and the elderly, requiring rapid and accurate diagnoses. This work develops a computational system to support the diagnosis of pneumonia in chest radiographs using Deep Learning. Eight pre-trained convolutional neural network architectures were evaluated: Xception, VGG16, VGG19, ResNet50, InceptionV3, InceptionResNetV2, DenseNet121, and DenseNet169, applying Transfer Learning and Fine-Tuning with different combinations of optimizers, learning rates, and batch sizes on the Chest X-Ray Images dataset. The F1-score was adopted as the main metric due to class imbalance, providing a more balanced evaluation. InceptionResNetV2 trained with SGD achieved the best performance, reaching an F1-score of 96%. A web interface was implemented using Streamlit. The results demonstrate that the solution can assist professionals in the early detection of pneumonia.