The effect of synthetic minority oversampling for enhancing brain tumor image classification

Brain tumors are among the most serious neurological diseases, posing significant diagnostic challenges due to their diverse nature and complexity imbalance in medical imaging datasets. To address these challenges, machine learning (ML) has demonstrated high potential in brain tumor classification; nevertheless, its overall performance can suffer when minority classes are underrepresented. This study examines the impact of the Synthetic Minority Over-sampling Technique (SMOTE) on MRI brain tumor classification using handcrafted features, including Gray Level Co-occurence Matrix (GLCM), Histogram of Oriented Gradients (HOG), and their combination (HOG + GLCM), with three classifiers, namely, Logistic Regression (LR), Support Vector Machines (SVM), and k-Nearest Neighbours (KNN). Experiments were carried out using two publicly available datasets.  For Dataset 1, SVM with GLCM increased from 55.21 to 57.40% (full SMOTE), but LR with HOG + GLCM increased from 52.4 to 53.49%. For Dataset 2, fold-wise SMOTE increased LR with GLCM from 69.85 to 70.60% and SVM with HOG increased from 92.87 to 93.10% compared to complete SMOTE. The results show that the SMOTE's effect is dependent on feature type, classifier, and augmentation strategy, with fold-wise typically boosting generalization while avoiding information leaking. These findings validate SMOTE as a viable method for improving the type overall performance in imbalanced medical imaging tasks, particularly for weaker texture-based descriptors.