A Novel Frequency-Domain Feature Extraction Framework for Machine Vision Applications

Abstract

This research addresses the challenge of brain tumor classification in MRI, specifically distinguishing low-grade gliomas (LGG) from high-grade gliomas (HGG). It presents a frequency-domain feature extraction framework that incorporates advanced preprocessing, optimization-based segmentation, and deep learning to enhance classification accuracy. Key stages include: mean filtering for noise reduction, Kapur's multilevel thresholding for tumor region segmentation, and a 2D Discrete Fourier Transform (DFT) for robust feature extraction. Optimized features are processed through a dual deep learning architecture, combining EfficientNetB3 for semantic feature extraction and Convolutional Neural Networks (CNNs) for preserving spatial hierarchies. The framework achieved over 95% classification accuracy on MRI datasets of LGG and HGG, demonstrating significant improvements over baseline methods. Its computational efficiency and innovative features offer a decision-support tool for clinicians in neuro-oncology, enhancing treatment planning and prognosis assessment.