For effective treatment options and better patient outcomes, early and accurate diagnosis of brain tumors is essential. This research introduces an innovative strategy to improving brain tumor diagnosis accuracy by combining deep learning with traditional machine learning classifiers. This research investigation employs the Xception Convolutional Neural Network (CNN) through a transfer learning approach as a feature extractor via two distinct strategies: (1) pruning the CNN’s classification layers while freezing the remaining layers, and (2) utilizing feature extraction with all CNN layers frozen. The extracted features are subsequently classified utilizing five traditional classifiers: Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), and Logistic Regression (LR). The suggested approaches are assessed using the BR35H: Brain Tumor Detection 2020 dataset, which is publicly accessible on Kaggle and includes a thorough collection of labeled MRI scans of the brain for both training and testing purposes. Results show that the hybrid models achieve exceptional performance, with both transfer learning-based strategies providing highly accurate tumor classification. Specifically, the Xception model with frozen CNN layers and feature extraction yielded testing accuracies of 0.9900 for Logistic Regression (LR) and 0.9850 for K-Nearest Neighbors (KNN). In comparison, pruning the CNN layers and freezing the remaining layers also resulted in comparable high performance, with testing accuracies of 0.9883 for KNN and 0.9900 for Logistic Regression (LR). According to these results, brain tumor diagnosis may be made much more efficient and accurate by combining deep learning feature extraction with standard machine learning classifiers.