Brain Tumor Classification in MRI Using Power Spectral Density and Spatial Autocorrelation Features

Abstract: Brain tumors represent a complex pathology characterized by abnormal cell growth in the brain and exhibit varying levels of aggressiveness depending on their tumor grade. Magnetic Resonance Imaging (MRI) is the reference modality for brain tumor diagnosis, providing detailed information about tissue morphology and heterogeneity. In this context, image-based feature analysis provides quantitative tools to extract informative descriptors for tumor characterization and classification.

This study proposes a feature-based methodology that exploits the space–frequency duality of MRI images to analyze tumor heterogeneity. Frequency-domain features are extracted from the Power Spectral Density (PSD), while spatial descriptors based on autocorrelation characterize the structural organization of tumor tissues. The combination of spectral and spatial features provides a multiscale representation of tumor heterogeneity.

Using Support Vector Machine SVM and Random Forest RF classifiers, PSD features achieve strong performance, with AUC values of 0.886, 0.904, 0.978, and 0.959 for SVM and 0.930, 0.957, 0.999, and 0.983 for RF across glioma, meningioma, no tumor, and pituitary classes. In contrast, autocorrelation features show lower performance with SVM, with an AUC of 0.861, but improve with RF, with an AUC of 0.947.

The fusion of PSD and autocorrelation features improves classification performance. Using SVM, AUC values reach 0.899, 0.905, 0.988, and 0.971, while RF achieves 0.927, 0.940, 0.997, and 0.984 for glioma, meningioma, no tumor, and pituitary classes. These results show that frequency-domain descriptors are highly discriminative, while combining spectral and spatial information improves the robustness of MRI tumor classification for computer-aided diagnosis.