Brain tumor classification from magnetic resonance imaging (MRI) plays a critical role in supporting radiologists during diagnosis and treatment planning. However, many existing automated approaches employ limited preprocessing, single-stage transfer learning, or evaluation on a single dataset, which restricts robustness and clinical applicability. This study proposes an enhanced transfer-learning framework based on the Xception architecture for multiclass brain tumor classification and compares its performance with baseline models under identical experimental conditions. The framework integrates a comprehensive preprocessing pipeline consisting of normalization, adaptive noise filtering, contrast enhancement, and targeted data augmentation, together with a structured two-phase fine-tuning strategy. A total of 6,537 MRI images were used, employing five-fold cross-validation, independent testing, and validation on an additional benchmark dataset. The proposed model achieved a mean cross-validation accuracy of 0.8994 ± 0.089 and 99.06% accuracy, precision, and recall on the independent test set, demonstrating strong stability and generalization ability. Evaluation on the CE-MRI Figshare dataset further confirmed robustness, yielding 98.45% accuracy, 98% precision, and 98% recall. In contrast, when re-evaluated within the same experimental setting, baseline models performed considerably worse: the SVM classifier achieved 21.41% accuracy, and ResNet50 reached 75.27%, both substantially inferior to Xception. Although higher accuracies for these models have been reported in prior studies under different conditions, the present findings highlight their limited generalization under unified evaluation. Overall, the proposed Xception-based framework provides a reliable and generalizable solution for automated brain tumor classification, with strong potential to support clinical workflows such as triage prioritization and second-opinion assistance.