Uma nova abordagem baseada em modelos do tipo encoder-decoder para segmentação semântica de lesões de fundo associadas à retinopatia diabética

Abstract

Diabetes is a chronic disease that affects the body’s ability to produce or use in sulin properly. Researchers estimate that approximately 537 million adults worldwide are affected by the disease, representing 10.5% of the population between 20 and 79 years old, and they project this number will reach 783 million by 2045. In this global context, Brazil occupies the fourth position in prevalence, with approximately 13.7 million cases. Diabetic Retinopathy (DR) is one of the main microvascular complica tions of diabetes and one of the leading causes of blindness in adults of working age. Since specialists usually diagnose DR by analyzing fundus images and identifying lesions such as Hard Exudates, Hemorrhages, Soft Exudates, and Microaneurysms, researchers can use deep learning approaches to segment these lesions. This work aims to propose a new approach to improve the accuracy in the segmentation of these lesions, which will contribute to a more effective diagnosis of the disease and facilitate the process of annotating images in new DR datasets. The proposed approach com bined the U-Net++, Attention U-Net, and R2U-Net architectures as ensembles. This work applied pre-processing, data augmentation, and pre-training steps to improve the extraction of lesion features. The experiments used two public RD datasets. The DDR dataset supported training, adjustment, and evaluation of the models composing the proposed approach, achieving an mDC of 63.75%, an mIoU of 83.18% in the validation stage, and an mDC of 66.47% and an mIoU of 81.15% in the testing stage. To evaluate the generalization capacity of the proposed approach on different DR datasets, this study employed the IDRiD dataset, reaching an mIoU of 84.36% in the validation stage and 82.84% in the testing stage. The results obtained were superior to those presented in related works, evidencing that integrating ensembles of encoder-decoder type models, combined with pre-processing, data augmentation, and pre-training steps, represents an effective strategy for segmenting fundus lesions.