Comparação entre Vision Transformers e U-Net com MobileNetV2 na segmentação de Exsudatos Lipídicos em imagens de retinografia

Abstract

Diabetic Retinopathy is a disease affecting people with Diabetes, which is among the leading causes of vision loss in individuals aged 20 to 75 in Brazil. With the rising number of Diabetes cases both nationally and globally, new technologies have emerged aiming to diagnose the early signs of the disease via fundoscopy examinations. Among the factors characterizing the disease, the presence of white lipid particles, termed lipid exudates, is common and visually evident in retinal images. This study investigated the performance of two approaches in the field of Computer Vision for the semantic segmentation of hard exudates, aiming to promote a comparative analysis between a classical approach and a state-of-the-art one optimized for mobile GPU, given the advancement in research and development of portable smartphone-based devices for fundus examinations. The U-Net model using MobileNetV2 as its architectural encoder, and MobileViT, a proposed Vision Transformer (ViT) focused on applications in mobile devices, were applied. Based on these two models, a comparison was performed using 3 metrics widely employed in Computer Vision semantic segmentation tasks – the Dice Coefficient, the Intersection over Union (IoU), and inference time, considering the Nvidia Tesla T4 GPU provided by Google Colab. MobileViT presented better results than U-Net, with a Dice coefficient surpassing the CNN-based model by 49%, in addition to demonstrating significant superiority in terms of computational efficiency. Although superior in metrics, the ViT model exhibited overfitting, even with the application of image augmentation techniques. The results evidence the computational superiority of MobileViT, pointing to this pre-trained model as a promising solution for the efficient semantic segmentation of medical images.