Classificação de doenças oftalmológicas em imagens retinianas com controle de confiança usando deep learning

Abstract

Ophthalmic diseases are among the leading causes of visual impairment and preventable blindness. This study proposed a reproducible pipeline composed of three independent and complementary modules: retinal vascular segmentation, multi-disease classification, and a clinical uncertainty mechanism. Each module was evaluated independently, without the output of one module serving as input to the next, using the FIVES dataset. Module 1, based on Attention U-Net with a ResNet50V2 encoder, performed vascular segmentation, achieving a Dice score of 0.8240 and an IoU of 0.7009, while providing overlay visualizations as interpretative support. Module 2, based on EfficientNetB3, classified fundus images into four clinical categories (Normal, Glaucoma, Age-related Macular Degeneration, and Diabetic Retinopathy), achieving an accuracy of 71.50%. Module 3 applied a clinical uncertainty rule based on confidence thresholds and the margin between classes, classifying 40 examinations (20.0%) as inconclusive and increasing the accuracy of accepted predictions to 78.12%, with a 38.6% reduction in errors. The results indicate that the combination of vascular segmentation, multi-disease classification, and confidence analysis brings the system's behavior closer to clinical practice, enabling the automatic identification of inconclusive cases for referral to specialized evaluation.