Challenging deep learning universality: a complete evaluation of conventional machine learning and deep learning in classifying brain tumors using magnetic resonance imaging

Background Accurate and timely classification is essential for brain tumor treatment planning. Classical machine learning (ML) with handcrafted features on moderate-sized datasets is under-represented, despite deep learning (DL) dominating medical image analysis. Purpose This study proposes a paradigm-agnostic comparison of classical ML and DL for brain tumor classification using a novel multi-domain feature framework, identifying which paradigm best balances accuracy, efficiency, and clinical use. Methods The Kaggle Brain Tumor magnetic resonance imaging (MRI) dataset (7,023 images; Glioma 1,621, Meningioma 1,645, No Tumor 2,000, Pituitary 1,757) was used. Nine feature domains—Gabor filters, local binary patterns (LBP), gray-level co-occurrence matrix (GLCM), discrete wavelet transform (DWT), discrete cosine transform (DCT), morphological, statistical, gradient, intensity gave 289 features, reduced to 150 by SelectKBest with the analysis of variance (ANOVA) F-test. Sixteen models were tested: 8 ML (Support Vector Machine with Radial Basis Function kernel (SVM-RBF), Random Forest, Gradient Boosting, Extra Trees, AdaBoost, K-Nearest Neighbors, Logistic Regression, Linear Discriminant Analysis) and 8 DL (Deep Multilayer Perceptron (Deep MLP), 1-Dimensional Convolutional Neural Network (1D-CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM), ResNet-inspired, DenseNet-inspired, Autoencoder). The Wilcoxon signed-rank test assessed significance. Shapley Additive explanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), single-domain, progressive, leave-one-out (LOO) ablations, data-size sensitivity, and ImageNet transfer-learning comparisons were performed. Results SVM-RBF achieved 96.80 ± 0.68% accuracy, 96.73% precision, 96.63% recall, 96.65% F1, 99.63% Receiver Operating Characteristic Area Under the Curve (ROC-AUC), Cohen’s Kappa 0.957, Matthews Correlation Coefficient 0.957, outperforming all 8 DL models ( p ≤ 0.0039). DenseNet- and ResNet-inspired, trained on the same handcrafted features, reached 94.86 ± 0.76% and 93.95 ± 0.82%. The ML mean (92.78 ± 3.06%) exceeded the DL mean (89.92 ± 4.50%, p = 0.0020), with shorter training (21.12s vs . 166.62s, p = 0.0020). Per-class recall ranged 93.2–99.0%. Gabor was the most influential domain (single-domain 94.85%; SHAP 0.122; LOO drop −1.25%, p < 0.001). Conclusions Classical ML with multi-domain handcrafted features outperformed all 8 DL architectures with shorter training time, challenging DL’s assumed universality in medical imaging. The framework also raised DL accuracy by 1–16 percentage points over raw-image transfer learning, showing engineered features benefit both paradigms. With strong interpretability (SHAP, LIME) and low computational needs, the approach suits clinical deployment in resource-limited settings.