Avaliação de aprendizado Auto-supervisionado na classificação de glomérulos renais em cenário de forte desbalanceamento de dados

Abstract

The diagnosis of chronic kidney diseases through histopathological analysis involves challenges such as inter-observer variability and the high time required for analysis. From a computational perspective, the automatic modeling of these lesions is further hindered by severe class imbalance and the scarcity of representative samples of rare lesions. This dissertation investigates the use of Self-Supervised Learning (SSL) techniques to build an automatic classifier of glomerular lesions, aiming to overcome limitations of supervised methods that depend on large amounts of labeled data. A total of 12,524 renal tissue images were used, distributed across eleven distinct lesions and obtained from multiple histological stains: hematoxylin and eosin (HE), periodic acid–Schiff (PAS), and Azan–Mallory (Azan). SSL methods such as SelfDistillation with No Labels (DINO) and Simple Framework for Contrastive Learning of Visual Representations (SimCLR) were evaluated with different backbones (ResNet and Vision Transformers), considering data augmentation and balancing strategies. The results demonstrate that SSL methods outperformed supervised approaches in multiclass classification tasks under imbalanced data scenarios. The DINO method, in particular, stood out in this context, achieving an F1-micro score of 58.27% compared to 43.47% for the ImageNet-pretrained baseline. In contrast, SimCLR showed superior performance in binary classification for several classes, highlighting its discriminative power in specific lesion classification tasks.