Brain Tumor Detection via HOG-FOS Feature Fusion with LDA and SMOTE: An Explainable AI Approach

Computer-aided diagnosis (CAD) systems have garnered significant attention for their potential to enhance clinical decision-making through faster and more accurate results, establishing them as a crucial research focus. The purpose of this work is to help radiologists improve patient care by introducing an effective CAD approach to distinguish between tumor and non-tumor categories in brain Magnetic Resonance Imaging (MRI) images. Our approach employs a novel feature fusion technique combining First Order Statistics (FOS) and Histogram of Oriented Gradients (HOG) to use both statistical and textural features. To address class imbalance, the SMOTE technique is used for dataset resampling. However, the extensive feature set may demand considerable computational resources. To mitigate this issue, feature dimensionality reduction is performed using the Linear Discriminant Analysis (LDA) technique. Classification is then conducted using boosting classifiers with five-fold crossvalidation. With a single feature, the suggested model attains 98.30\% accuracy and area under the curve is 1. The increasing importance of AI in the healthcare industry has sparked concerns about its potential bias, lack of transparency, and explainability. By including the SHapley Additive exPlanations (SHAP) value to evaluate the significance of attributes in the model's predictions, our method fixes these problems and enhances the system's interpretability.