<![CDATA[Brain tumor classification from magnetic resonance imaging (MRI) images requires robust feature learning because tumor classes show subtle visual differences, especially between glioma and meningioma. This work improves accuracy and stability in multi-class brain tumor classification by enriching features and directing attention to discriminative MRI patterns. A hybrid model uses DenseNet121 and MobileNetV2 as parallel backbones, applies a Convolutional Block Attention Module (CBAM) to each stream, and performs late fusion by concatenating the two feature vectors for final prediction. CBAM improves performance by adaptively reweighting channels that encode tumor-relevant intensity and texture cues while suppressing less informative responses. It also strengthens spatial focus by emphasizing regions aligned with tumor cores and boundaries, reducing interference from normal tissue and artifacts and lowering cross-class confusion. Performance is evaluated with 10-fold cross-validation on the Figshare dataset and a standard hold-out split on the Sartaj Bhuvaji dataset. The proposed model achieves strong results on Figshare (3 classes) with accuracy 0.9892, precision 0.9880, recall 0.9876, F1-score 0.9877, and AUC 0.9995, and it also achieves strong results on Sartaj Bhuvaji (4 classes) with accuracy 0.9816, precision 0.9814, recall 0.9848, F1-score 0.9829, and AUC 0.9984. These results indicate that dual-backbone fusion increases feature diversity and CBAM reduces cross-class confusion, supporting stable generalization across independent MRI datasets. The model improves accuracy in separating visually similar tumors, especially glioma versus meningioma, and maintains strong generalization across different datasets.]]>
Copyrights © 2026
