MULTI-PLANE ATTENTION-AIDED CNN FOR DETECTION OF BRAIN TUMOR IN MRI SCANS

Abstract

The classification of brain tumors from MRI images plays an important role in assisting radiologists to get accurate and reliaable diagnostic decisions. Existing deep learning approaches often rely on single-plane MRI analysis, which limits the exploitation of complementary spatial information available across multiple anatomical views. To address this limitation, we have proposed Multi-Plane Attention-based Convolutional Neural Network (MPA-CNN) for automatic detection of brain tumor. The proposed framework independently processes axial, coronal, and sagittal MRI planes using customized EfficientNet-B0 backbones to extract discriminative deep features. An attention-based fusion mechanism is then employed to adaptively weight and integrate multi-plane features into a unified representation, which is subsequently classified using a fully connected prediction head with label smoothing. The proposed model is analyzed on BRISC2025 brain MRI dataset which contain four classes. Among these classes, one is healthy class named ‘no_tumor’ and remaining three classes are diseased one, which are: glioma, meningioma, and pituitary. The experimentatl findings shows that our proposed model has achieved precision of 99.31%. It is achieved 99.30% for precision, recall, and F1-score. Moreover, it has achieved highest class-wise performance, and ROC-AUC values for all types of tumors. On the basis of these results, it is concluded that multi-plan attention fusion for automated brain tumor classification is very effective and reliable in clinical decision-making. These results confirm the effectiveness and robustness of multi-plane attention fusion for automated brain tumor classification, highlighting its potential for reliable clinical decision support.