Avaliação de arquiteturas de deep learning para esclerose múltipla: de baselines convolucionais a redes híbridas vit e prototípicas

Abstract

The diagnosis of multiple sclerosis (MS) by magnetic resonance imaging (MRI) remains a challenge due to the complex spatial distribution of the lesions. Although Convolutional Neural Networks (CNNs) are effective at identifying local textures, they often fail to capture the broader spatial relationships between distant areas of the brain. In this study, we propose and compare three deep learning architectures for automated MS classification: a standalone ResNet18 baseline, a hybrid CNN-ViT model trained with cross-entropy loss, and a CNN-ViT variant trained with prototypical loss for metric learning. To ensure a reliable performance assessment, all models were validated using a 5-fold cross-validation and an independent test set of 60 subjects. The hybrid CNN-ViT architecture achieved 86.67% accuracy and 90.00% specificity, consistently outperforming the CNN-only baseline. The prototypical variant achieved 83.33% accuracy and 80.00% specificity, offering a complementary metric learning perspective. Additionally, we analyzed the model’s attention maps to interpret its decision-making process, which highlighted both its ability to focus on relevant neurological features and its current limitations with non-brain tissues. These results indicate that combining self-attention mechanisms with convolutional layers improves classification performance and interpretability, providing a robust framework for automated MS detection even with limited datasets.