FlexiCombFE: A flexible, combination-based feature engineering framework for brain tumor detection

Abstract

Deep learning models are prevalently employed for image classification, but significant opportunities remain within the domain of feature engineering. This study introduces FlexiCombFE, a novel, flexible, combinationbased feature engineering framework for brain tumor detection using various fixed-size patch divisions. This framework employs three distinct feature extractors (Local Phase Quantization, Local Binary Pattern, and Pyramidal Histogram of Oriented Gradients) to generate a total of seven primary feature vectors. By using four types of fixed-size patch divisions, 28 feature vectors are generated. Then, three feature selectors (Chi-squared, Neighborhood Component Analysis, and ReliefF) create 84 selected feature vectors. In the classification phase Knearest neighbors and support vector machine classifiers yield 168 classifier-specific outcomes. An information fusion generates 166-voted outcomes, with the most accurate classification outcome selected as the final output. This self-organizing feature engineering model achieved a classification accuracy of 99.35% on a brain tumor image dataset, outperforming several deep learning approaches. The modular design of the proposed framework allows for detailed analysis of the classification effects of individual methods, providing optimal feature engineering strategies for medical image classification tasks.