Attention Enhanced Deep Learning for MRI-Based Brain Tumor Classification: A Comparative Ablation Study

Objective: The performance of models was evaluated using an ablation-based approach employing attention mechanisms and transfer learning for tumor classification in brain magnetic resonance imaging images. Methods: The glioma, meningioma, no tumor, and pituitary classes from the BRISC2025 public dataset were used. ResNet50, DenseNet121, EfficientNetB0, Xception, and InceptionResNetV2 models were tested using a baseline and the squeeze-and-excitation (SE) attention, spatial attention, and convolutional block attention module attention constructs. The models were trained in two phases: frozen backbone training and fine-tuning, with each experiment repeated three times. Performance was evaluated using accuracy, precision, recall, F1-score, area under the curve (AUC), and loss values. Results: The most successful model was EfficientNetB0 + SE Attention. This model achieved an accuracy of 0.9580 ± 0.0062, an F1-score of 0.9580 ± 0.0057, and an AUC of 0.9949 ± 0.0012 on the test set. Class-based results showed particularly high success rates for the no-tumor and pituitary classes. The most challenging class was meningioma. Conclusion: Transfer learning models supported by attention mechanisms have yielded successful results in brain tumor magnetic resonance classification. The EfficientNetB0 + SE Attention model stood out as the most balanced and successful architecture among the evaluated methods. © 2026 The Author(s). Published by Galenos Publishing House on behalf of the Turkish Society of Magnetic Resonance.