Neuro-Oncology Benchmark: The Resource-Interpretability Tradeoff in Radiomics and Multiclass Brain Tumor Classification Based on Deep Transfer Learning vs Handcrafted Radiomics.

Othmane Bakkas, Drissia Ennagoura, Nasreen Badruddin, Abdelali Zbakh, Mohamed El Mahjouby, Badre Bossoufi, Khalid El Fahssi, Mohamed El Far, Mohamed Taj Bennani

The automatic multiclass brain tumor classification using MRI images plays an important role in a non-invasive clinical setting. However, the choice of a model demands the trade-off between accuracy, complexity, and interpretability of the classifier. In this study, we have established a benchmark for comparison of classical machine learning (ML) approaches to deep learning (DL) methods, using the same data distribution setup to describe the aforementioned complexity-interpretability trade-off frontier. We generate a 72-dimensional manually designed radiomic feature set, which comprises the first-order intensity features, GLCM features, LBP features, shape morphological measures, and FFT characteristics. These are compared against fine-tuned, ImageNet-pretrained ResNet50 and EfficientNet-B0 on a balanced 4-class dataset (Glioma, Meningioma, Pituitary, and No Tumor; $n=7,200$ images). To address the clinical "black-box" problem, a dual-axis Explainable AI (XAI) pipeline maps SHAP values to the radiomic feature space and Grad-CAM activations to the deep neural layers. Results show that, based on three independently seeded training cycles, ResNet50 (95.44% $\pm$ 0.28%) and EfficientNet-B0 (95.44% $\pm$ 0.22%) achieve statistically indistinguishable accuracy, while the radiomic-driven SVM reaches a robust 89.12% accuracy and 0.9613 Macro AUC, training in under 4 seconds versus over 17 minutes for EfficientNet-B0. Five-fold cross-validation confirms the stability of the SVM pipeline (92.43% $\pm$ 0.89%). Despite matching ResNet50's accuracy, EfficientNet-B0 achieves this while reducing parameters by 83% (4.0M vs. 23.5M) and training in roughly half the time. We conclude that radiomic pipelines suit resource-constrained edge deployment, while lightweight deep networks integrated with visual XAI tools offer the ideal configuration for centralized diagnostic frameworks.