Título: Comparative Analysis of the Impact of Data Augmentation Techniques on Different Encoders for Diabetic Retinopathy Segmentation
Autores: Maria Eduarda Ribeiro Donato da Silva, Leonardo Vidal Batista & Túlio Emanuel Santana de Souza
Resumo: This work presents a systematic analysis of data augmentation techniques applied to diabetic retinopathy segmentation in retinal images, with a focus on hard exudates. Five encoders with the UNet++ architecture were used: MobileNet-v2, VGG11, VGG16, ResNet50, and EfficientNet-B3, evaluated on the IDRiD dataset. The data augmentation techniques analyzed include horizontal flip, shift scale rotate, RGB shift, random brightness contrast, and elastic transform. The metrics used were Intersection over Union (IoU) and Dice Coefficient. The results indicate that complex combinations of techniques do not always lead to performance improvements. The best configuration was the combination of horizontal flip, elastic transform and random brightness contrast, which achieved a Dice of 0.676 and IoU of 0.510 with the VGG11 encoder. However, horizontal flip achieved highly competitive metrics when applied in isolation. The experiments highlight the importance of evaluating each augmentation technique’s impact both individually and in combination in medical segmentation tasks.
Palavras-chave: Data Augmentation; Diabetic Retinopathy; Semantic Segmentation; Retinal Images; UNet++.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175473
Artigo em PDF: CBIC_2025_paper1175473.pdf
Arquivo BibTeX:
CBIC_2025_1175473.bib