Comparative Evaluation of Ensemble Models for Brain Tumor Image Classification Using Wavelet Transform-Based Features

Abstract

This study aims to develop and evaluate the performance of machine learning models for classifying magnetic resonance (MRI) brain images with tumor conditions. The process began with the extraction of multi-level texture and frequency features using the Discrete Wavelet Transform (DWT) with Daubechies 4 as the mother wavelet. The decomposition produced a total of ten subbands, including one approximation subband (LL), three horizontal detail subbands (LH), three vertical detail subbands (HL), and three diagonal detail subbands (HH). Each subband was characterized by six statistical parameters: mean, standard deviation, energy, entropy, skewness, and kurtosis, which effectively represented the structural characteristics of brain tissues. In total, 60 features were obtained. Feature selection was then performed using the Random Forest algorithm based on feature importance scores, where only features with importance values greater than 0.01 were retained to enhance accuracy and reduce data dimensionality. This process resulted in 39 selected wavelet features from the all original ones, corresponding to a 35% reduction in unnecessary attributes. For model construction, four ensemble learning techniques were implemented, Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Random Forest, and Bootstrap Aggregating (Bagging), to compare classification performance. Experimental results revealed that the AdaBoost model achieved the best performance, with an accuracy of 93.67%, sensitivity of 94.78%, specificity of 93.00%, and overall performance of 93.81%. These findings reflect that the integration of wavelet transform-based feature extraction and feature selection with ensemble learning methods, particularly AdaBoost, can effectively and accurately support medical diagnosis from brain MRI images, providing an additional approach for brain tumor detection.