Explainable brain tumor classification using an attention-augmented CNN and Grad-CAM

Creative Commons License

Mustafa M. A., Erdem O. A., Söğüt E.

Journal of Computer & Electrical and Electronics Engineering Sciences, cilt.4, sa.2, ss.77-82, 2026 (Hakemli Dergi)

Özet

Aims: This paper explores clinical explainability of deep neural network architectures for classifying brain neoplasms using magnetic resonance imaging (MRI) scans.

Methods: We evaluated multiple architectures, including a baseline convolutional neural network (CNN), a novel attention- augmented CNN, and pre-trained transfer learning models (MobileNetV2, InceptionV3, and Xception). Models were trained using the Adam optimizer on a quaternary-class MRI dataset comprising glioma, meningioma, pituitary adenoma, and healthy cases. Performance was assessed using quantitative metrics (accuracy, precision, sensitivity, and F1-score), while interpretability was evaluated via gradient-weighted class activation mapping (Grad-CAM) visualizations.

Results: MobileNetV2 demonstrated superior performance, achieving a classification accuracy of 91.7%, with highly consistent precision, sensitivity, and F1-scores. The attention-augmented CNN achieved competitive efficacy, while Grad-CAM heatmaps consistently localized pathologically significant regions, thereby confirming the diagnostic reliability and transparency of the models’ decision-making processes.

Conclusion: Computationally efficient pre-trained architectures like MobileNetV2 are highly effective for MRI-based neoplasm classification. Furthermore, integrating attention mechanisms with Grad-CAM visualization significantly enhances both diagnostic accuracy and model transparency, highlighting their strong potential for adoption in real-world clinical decision support systems.