Implementação e avaliação de redes neurais compactas para detecção de catarata com dados limitados

Abstract

Cataract is one of the leading causes of vision loss worldwide, and early detection plays a key role in timely medical intervention. In this context, neural networks have stood out as powerful tools for analyzing complex images and extracting relevant information for accurate diagnoses. The objective of this work is to implement and evaluate the effectiveness of compact neural networks in the automatic detection of cataracts in eye images using deep learning, in a scenario with a limited database, both in terms of the number of images and their resolutions. It is initially carried out a literature review of existing techniques and approaches used in the area. The themes and concepts used to carry out the study are also reviewed. The process includes building appropriate neural network architectures, preprocessing medical fundus images to improve data quality, balancing the database, implementing deep learning algorithms for model training, and then evaluating performance using metrics such as accuracy, sensitivity, and loss. The results obtained demonstrate that models based on ResNet architectures, both in their traditional versions and hybrid versions with Vision Transformers (ViT), presented high performance (above 93%) in cataract detection. On the other hand, compact models such as the Compact Convolutional Transformer (CCT) and the Compact Vision Transformer (CVT) faced difficulties in generalizing to multiple classes, suggesting that compact transformers.