Uso de Few-shot Learning para lidar com o desbalanceamento severo de dados na classifica??o de les?es glomerulares

In the field of nephropathology, the automatic classification of renal biopsy images is crucial for supporting patient diagnosis and treatment. Machine Learning (ML) techniques?particularly those based on Convolutional Neural Networks (CNNs)? have enabled the automation of this task, offering operat...

Descripción completa

Detalles Bibliográficos
Autor: Santos, Alexsandro Silva
Tipo de recurso: tesis de maestría
Estado:Versión publicada
Fecha de publicación:2025
País:Brasil
Institución:Universidade Estadual de Feira de Santana (UEFS)
Repositorio:Biblioteca Digital de Teses e Dissertações da UEFS
Idioma:portugués
OAI Identifier:oai:tede2.uefs.br:8080:tede/1929
Acceso en línea:http://tede2.uefs.br:8080/handle/tede/1929
Access Level:acceso abierto
Palabra clave:Patologia Computacional
Desbalanceamento de classes
Few-shot Learning
Computational Pathology
Class Imbalance
CIENCIAS EXATAS E DA TERRA
Descripción
Sumario:In the field of nephropathology, the automatic classification of renal biopsy images is crucial for supporting patient diagnosis and treatment. Machine Learning (ML) techniques?particularly those based on Convolutional Neural Networks (CNNs)? have enabled the automation of this task, offering operational improvements. However, these techniques typically require large, balanced datasets to achieve strong performance and are negatively impacted in scenarios with limited data and class imbalance?conditions commonly found in the classification of rare renal lesions. This study presents an applied investigation of approaches for handling class imbalance, including well-established techniques from the literature and the use of Few-Shot Learning (FSL). The results demonstrate that it is possible to achieve F1- scores above 85% across different imbalance scenarios encountered in the analyzed lesion types, reaching scores above 90% even in cases with imbalance ratios as high as ? 1:30. Nevertheless, the experiments conducted in this study revealed that data imbalance, when considered in isolation, is not the sole factor influencing the final performance of ML models.