Predictive Algorithm for Brain Tumor Detection Using Explainable Artificial Intelligence: A Radiomics-Based Machine Learning Approach
Keywords:
Brain Tumor Detection; Explainable Artificial Intelligence (XAI); SHAP; Grad-CAM; Machine Learning; MRI Classification; Radiomics; Clinical Decision SupportAbstract
Brain tumors remain among the most challenging conditions in neuro-oncology because early, accurate diagnosis directly determines treatment options and survival outcomes, yet conventional deep-learning classifiers are frequently criticised as opaque ‘black boxes’ that clinicians cannot audit. This paper proposes and empirically evaluates a predictive, explainable machine-learning framework for classifying brain magnetic resonance imaging (MRI) slices as tumorous or non-tumorous. The system pipeline integrates pre-processing, radiomic-style feature engineering (statistical, textural, and asymmetry-based descriptors), a benchmarked ensemble of five predictive models (Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, and a Multi-Layer Perceptron), and two complementary explainable artificial intelligence (XAI) layers: SHapley Additive exPlanations (SHAP) for global and local feature attribution, and an occlusion-based pixel saliency method that localises the image regions driving each prediction. On a held-out test partition, the best-performing model (MLP) achieved 98.7% accuracy, 98.6% F1-score, and an AUC of 0.998, while the designated explainable reference model (Random Forest) reached 96.0% accuracy with fully transparent, exact Shapley-value attributions. SHAP analysis identified intensity variance, low-intensity area fraction, image entropy, and hemispheric asymmetry as the dominant predictive signals — consistent with established radiological markers of tumor presence — while saliency maps confirmed that the models attend to the lesion region itself rather than to spurious background cues. These results demonstrate that interpretable, feature-based predictive pipelines can approach the accuracy of opaque deep networks while providing the transparency required for clinical trust, auditability, and regulatory acceptance.