Uma nova abordagem para redes neurais convolucionais dinâmicas para sistemas CAD em ambiente federado

Abstract

In the context of Artificial Intelligence, Convolutional Neural Networks (CNNs) present themselves as an alternative for identifying patterns in images, which are often not visible to the naked eye. With the advancement of machine learning techniques that demand increasing computational power and large amounts of data to train these techniques, computational resources become a limitation. To address this issue, dynamic convolutional networks that modify their architectures have been implemented to deal with devices with memory and processing constraints. This thesis presents a new approach to dynamic convolutional networks that adapt their architecture and weights according to different types of input, aiming to improve efficiency and accuracy in pattern identification in biomedical images. The VGG-16 convolutional network was used as the basis for implementing the proposed approach. The CIFAR-10 dataset was used for comparative reference, and five glaucoma image datasets were used to evaluate the performance of the proposed approach based on biomedical images. The metrics applied to evaluate the implemented models are: accuracy, precision, recall, F1-score, and the number of network parameters. Additionally, the Wilcoxon test is used to assess statistical differences between the results of the original and dynamic architectures. For the CIFAR-10 dataset, the proposed approach reduces the size of the original VGG-16 network by up to half, with a maximum loss of 7% in accuracy, precision, recall, and F1-score. For most of the datasets, the reduced versions perform equivalently or better compared to the original network, particularly in the external validation set, which better represents real-world situations in diagnostic support systems. Finally, the proposed approach is implemented in a federated environment, aiming to perform decentralized training in hospitals while ensuring data privacy and greater robustness in the final model. The results obtained with the CIFAR-10 dataset and six glaucoma image datasets show that the proposed approach reduces the size of the original VGG-16 network by up to half, with minimal losses in accuracy, precision, recall, and F1-score metrics, making it possible to use the proposed methodology in federated environments with processing constraints. The main contribution of this thesis lies is the new approach developed, which can be applied to any convolutional neural network, transforming them into dynamic networks.