The automatic classification of brain tumors in magnetic resonance imaging (MRI) remains a challenging task due to the variability of tumor morphology and the heterogeneity of available datasets. This work presents a comparative evaluation of three deep learning models (ResNet-50, DenseNet-121, and ViT-B16) using the recently curated BRISC 2025 dataset and the BCD-MRI dataset. A systematic evaluation protocol was adopted, including internal validation, external evaluation restricted to test splits, and external evaluation over the entire external datasets. Four levels of data augmentation, ranging from none to strong, were applied to assess their impact on generalization. Internal evaluations achieved high and stable results across all models, while external test set evaluations showed only minor drops. More pronounced differences emerged when models were tested on entire external datasets: knowledge transferred from BCD-MRI generalized better to BRISC than the reverse. Data augmentation reduced part of the performance loss, particularly in external evaluations, yielding significant gains in robustness. Overall, DenseNet-121 proved to be the most reliable architecture, ResNet-50 showed resilience when combined with augmentation, and ViT-B16 achieved strong results in restricted cases but was less stable in whole-dataset external evaluations. The main contribution of this study is to demonstrate that progressive data augmentation enhances the external dataset generalization, highlighting the importance of evaluating models beyond internal splits or limited external test sets to better approximate real-world deployment conditions.
Viana et al. (Tue,) studied this question.