Desenvolvimento de Abordagens Baseadas em Redes Neurais Profundas para Detecção e Segmentação de Instância de Lesões Retinianas

Abstract

Diabetic Retinopathy (DR) is one of the leading causes of vision loss and presents fundus lesions in its initial stages, such as microaneurysms, hemorrhages, hard exudates, and soft exudates. Computational models capable of detecting these lesions can support the early diagnosis of the disease and prevent the manifestation of more severe forms of lesions, helping in the screening process and definition of the best form of treatment. However, the detection of microlesions using computational systems is a challenge due to several factors, such as the size and shape of these lesions, the presence of noise and poor contrast in the images, the small number of labeled examples in public DR datasets, and the difficulty of deep learning algorithms in detecting tiny objects due to gradient dissipation during training. Thus, to overcome these problems, this work proposes two new approaches based on image processing techniques, data augmentation, transfer learning, and deep neural networks to support the medical diagnosis of fundus lesions. We trained, adjusted, and evaluated the proposed approaches using different public Diabetic Retinopathy datasets. We partitioned the datasets into sets of training (50%), validation (20%), and test (30%) to carry out the experiments. We used a validation step to fine-tune the hyperparameters and a test step to assess the generalization capacity of the models. The approach to detecting fundus lesions achieved mAP of 0.2630 for the limit of IoU of 0.5 in the validation step using the DDR dataset and Adam optimizer. The approach for segmenting instances of fundus lesions reached mAP of 0.2903 for the limit of IoU of 0.5 in the validation stage using the DDR dataset and Adam optimizer, thus being 10.38% more accurate than the proposed detection approach. The results obtained in the experiments demonstrate that the new approaches presented promising results in detecting fundus lesions associated with DR.