Background Automated classification of brain tumors from magnetic resonance imaging (MRI) can support radiologists and accelerate treatment planning. Public benchmark datasets enable rapid prototyping but require rigorous evaluation and transparent reporting. Methods The publicly available Kaggle Brain Tumor MRI dataset (DOI: 10.34740/kaggle/dsv/2645886 ) comprising 7,023 contrast-enhanced, T1-weighted axial slices labeled as glioma, meningioma, pituitary tumor or no-tumor was used. After removing corrupted images and applying extensive augmentation, a convolutional neural network was trained via transfer learning. A Residual Network 50 (ResNet50) backbone pretrained on ImageNet was fine-tuned in a three-phase schedule: (i) frozen feature training of custom classifier layers, (ii) partial unfreezing with a reduced learning rate and (iii) full fine-tuning of all layers. Regularization strategies included dropout, Gaussian noise, L2 weight decay and label smoothing. Performance was evaluated on a held-out test set ( n = 1,205) using accuracy, precision, recall, F1-score and confusion matrix analysis. Model interpretability was assessed with Grad-CAM++ heatmaps. Results The proposed model achieved 96.67% overall accuracy and a macro-averaged F1-score of 0.9658 on the unseen test set. Per-class recall ranged from 0.94 (meningioma) to 0.99 (pituitary). Training and validation curves indicated minimal overfitting, while Grad-CAM++ visualizations suggested that salient regions generally corresponded to tumor locations rather than background artifacts. Discussion These results demonstrate that a carefully regularized, fine-tuned ResNet50 provides a strong baseline for four-class brain tumor classification. Limitations Despite aggregating three subdatasets, the Kaggle corpus remains limited in diversity: all images correspond to axial, contrast-enhanced T1-weighted scans, and metadata on scanner type or acquisition protocol is unavailable. However, external clinical validation was not performed. Future work Directions include evaluation under domain-shift, validation on multi-institutional cohorts, extension to volumetric 3D models, and exploration of lightweight architectures for real-time deployment. Conclusions This work provides a reproducible baseline for interpretable brain tumor MRI classification, highlighting both the promise and current limitations of deep learning approaches prior to clinical validation.
Pietro Veneri (Thu,) studied this question.