The accurate classification of brain tumors based on magnetic resonance imaging (MRI) remains a persistent problem due to high intra-class diversity and subtle inter-class disparity in tumor shape. In this work, we introduce an innovative ensemble framework for tumor classification from axial T1-weighted contrast-enhanced MRI by integrating heterogeneous feature representations using a meta-learning-based intermediate fusion approach. The architecture consists of three encoding branches: a transfer learning path relying on a pre-trained MobileNetV2 for encoding anatomical priors, a residual dilated encoder for spatial continuity and long-range dependencies capture, and a multi-scale inception-based branch to gather features at different receptive fields. Instead of early or late fusion, the model performs intermediate fusion where the outputs from all three branches are concatenated and then further refined with a spatial attention-based meta-learner. This module re-weights the importance of features and improves discriminative representation before classification. A compact dense module is adopted to generate the final prediction for multi-class tumor subtype labeling. The model obtained classification accuracies of 96.4 % and 98.42 %, respectively, on Masoud Nickparvar’s and Cheng-Jun brain tumor datasets, while having an AUC greater than 99 % for all tumor classes in both cases. The results demonstrate the effectiveness of ensemble fusion along with attention-based meta-learning for representing complex tumor characteristics, thereby providing a promising direction for generalizable brain tumor diagnosis.