Deep learning aplicado ao diagnóstico da degeneração macular relacionada à idade

Abstract

Age-Related Macular Degeneration represents one of the main causes of irreversible blindness in the elderly, and its early diagnosis in Brazil faces significant challenges due to the unequal distribution of medical specialists. This work's main objective was to develop and evaluate an artificial intelligence system to act as a screening tool for the disease, using retinography images, which is a more accessible exam. The methodology adopted a quantitative and experimental approach, comparing the performance of two convolutional neural network architectures, DenseNet-121 and EfficientNet-B1. The models were trained with transfer learning and data augmentation techniques and validated through cross-validation. The results indicated that the EfficientNet-B1 model, despite a slightly lower overall accuracy, was superior in the sensitivity (recall) metric, achieving 72.58%. This metric is clinically more relevant as it demonstrates a greater ability to correctly identify positive cases, minimizing the risk of missed diagnoses. It is concluded that the developed system is a pragmatic and effective screening tool, with high potential to optimize the workflow of specialists and democratize access to eye health, especially in remote regions.