<![CDATA[Clinical evaluation of brain magnetic resonance imaging is challenged by subtle intensity variations and heterogeneous tumor boundaries. Many existing classification studies rely on shared repositories without explicitly addressing data redundancy, which may lead to biased performance estimation. This study introduces a brain tumor classification framework built upon an ECA-ResNet50 architecture to improve predictive consistency while maintaining computational efficiency. The ResNet50 backbone is augmented with an Efficient Channel Attention mechanism to enhance channel dependency modeling without substantially increasing model complexity. To preserve data integrity and reduce potential leakage, duplicate images were identified and removed using an MD5-based hashing procedure prior to model development. The predefined training and testing partitions supplied with the dataset were retained for evaluation to ensure non-overlapping samples. Experimental results demonstrate an overall accuracy of 90.79% and a recall of 77% for meningioma classification. Grad-CAM visualization indicates that model predictions are primarily associated with clinically relevant tumor regions, supporting interpretability and alignment with diagnostic structures. These findings suggest that combining lightweight attention mechanisms with systematic data curation enhances reliability and transparency in MRI-based brain tumor classification.]]>
Copyrights © 2026
