A Deep learning framework for brain tumor detection using CNNs and transfer learning on MRI scans

Abstract
Precise detection of brain tumors from magnetic resonance imaging is an important aspect of medical diagnostics, necessitating methods combining precision and robustness. This study presents the deep learning and transfer learning methods for brain tumor classification. CNNs, in addition, transfer learning techniques are used to classify MRI images into four groups: pituitary tumor, meningioma, glioma, and no tumor. To perform experimental analysis, this research employs an ensemble CNN model and pre-trained models, like MobileNetV2, Vision Transformer (ViT), and VGG16. The experimental outcomes show that the ensemble CNN technique had accomplished an average training and testing accuracy of 95.61 % and 96.72 % with balanced F1-score, precision, and recall of around 96 % and above. The ViT transfer learning algorithm established an average training and testing accuracy of 98.42 % and 96.72 % with F1-score, precision, and recall of around 96 % and above. The MobileNetV2 algorithm had achieved an average train and test accuracy of 99.80 % and 97.48 % with balanced F1-score, precision, and recall of about 97 % and above. Remarkably, the VGG16 algorithm further improved performance with an average train and test accuracy of 99.36 % and 98.78 % with balanced precision, recall, and F1 metrics of about 98 % and above. All the anticipated models\' specificity results continuously exceeded 0.98, emphasizing the model\'s consistency. These results exhibit how well ensemble methods and transfer learning perform for MRI-based tumor categorization. VGG16 exhibited the preeminent balance of accuracy, efficiency, and generalization amongst the algorithms assessed, demonstrating that VGG16 can be employed in clinical environments to support radiologists in analyzing brain tumors precisely.

Author
Sarhang Hayyas Mohammed

DOI
https://doi.org/10.1016/j.sasc.2025.200389

ISSN
2772-9419

Publish Date: 17-Sep-2025