Implementação de um modelo de deep learning para detecção de retinopatia diabética aplicado em uma base de imagens de retina

Abstract

The following study is based on the development of automated techniques that are capable of identifying the presence of diabetic retinopathy in retinal images. For this, deep learning was used, a subfield of study in the area of ​​artificial intelligence, where 5 (five) models were trained using the ResNet18 convolutional neural network architecture. Of these five models, only one focused on detecting diseased eyes at any stage of the disease, while the other four were trained to identify at each stage of the disease. At the end of training the models that detected the disease by stage, a pipeline was set up, with the four models that detected by stage, so that it could identify whether or not the retinal image showed signs of the disease. All models showed an accuracy above 90%, however, the model that compared healthy eyes with eyes with the initial stage of the disease, the specificity of this model was lower than the others, indicating few presence of disease characteristics in the image. For future work, we suggest increasing the data set and training models with other architectures on the same set.