Avaliação da classificação multiclasse de tumores cerebrais em ressonância magnética com transfer learning e redes convolucionais utilizando o banco de dados Brain MRI ND-5

Abstract

This study investigates the use of deep learning models for multiclass classification of brain tumors in magnetic resonance imaging (MRI), considering four clinically relevant categories: glioma, meningioma, pituitary tumor, and no-tumor. The architectures ResNet50, DenseNet121, EfficientNet-B5, and EfficientNet-V2-S were evaluated under a transfer learning framework with supervised fine-tuning. Experiments were conducted on two widely referenced datasets, Brain MRI ND-5 and Brain Tumor Classification (Bhuvaji et al.), in order to assess not only the absolute performance of the models but also their generalization capacity across different imaging domains. The methodology included dataset standardization, uniform training procedures, quantitative evaluation using classification metrics, analysis of learning curves, and interpretability assessment using Gradient-weighted Class Activation Mapping (Grad-CAM). Results showed high performance across all architectures in the ND-5 dataset, with accuracy around 97% and F1-scores and AUC-ROC values above 0.96 for all classes. In the Brain MRI dataset, performance varied among models: DenseNet121, EfficientNet-B5, and ResNet50 achieved accuracy above 98%, whereas EfficientNet-V2-S exhibited a notable decrease, indicating stronger sensitivity to domain shift. Confusion matrices revealed that misclassifications mostly occurred between glioma and meningioma, tumor types known for their radiological similarity. Interpretability analysis indicated that, in correctly classified cases, the models concentrated attention on tumor regions, while misclassified images displayed dispersed or anatomically inconsistent activation patterns. The findings demonstrate that pretrained convolutional neural networks can effectively classify brain tumors in MRI scans with high discriminative capability and spatially coherent decision patterns, reinforcing their potential for clinical decision-support systems. At the same time, the results highlight challenges related to dataset variability, suggesting the need for future research involving 3D volumetric approaches, multimodal integration, and validation in real clinical environments.