PREDICT STAGE OF DIABETIC RETINOPATHY USING DEEP LEARNING

Resumo

The diabetic retinopathy (DR) is one of main complications of diabetes mellitus (DM),
presenting in about 40% of diabetics and being the leading cause of blindness in 16 to 64 years old
people.The DR is caused by damage to the blood vessels. The diagnosis is given through analysis of
images generated by retinoscopy, that is an easy operating electronic device. Despite this, in the
Brazilian National Unified Health System (SUS), the access to this type of medical assessment
specialized is still poor, resulting in long queues until an ophthalmological consultation. The latest
Brazilian Diabetic Society Guideline recommends annual evaluations for DR in diabetic patients, a
non-realistic scenario for SUS patients. In view of this, a screening tool that could predict the RD and
that could be used by primary care practitioners could be an awesome public health solution. Then, we
presenting a deep learning framework (DLF) to classify DR stages. In this paper, we will perform the
entire DLF cycle: data collection, data preparation, model creation and validation. The main novelty is
the use of siamese convolutional neural network (SCNN), which will receive input pairs of eye fundus
images.The purpose of using this set of tools is to use the layers to extract the main characteristics of
the inputs of each neural network, so the weights of layers are shared and the similarity degree
between neural networks outputs are measure. We train this network using a high-end graphics
processor unit (GPU) on the publicly available Messidor dataset and, with this approach, we expect
better results in predicting the DR stage when compared to others works.