Fast texture-based analysis and classification of 3D MRI brain scans

Texture-based analysis of medical images is an increasingly important diagnostic tool. This paper presents a method for the construction and application of two-and three-dimensional Zernike polynomial-based translationally-rotationally-mirroring (TRM) image texture invariants, derived via the Fourier transform, for image texture analysis and classification tasks. We propose an approach for computing global characteristics from these invariants, enabling fast and efficient feature extraction. The method is evaluated on a binary classification task that distinguishes between cognitively normal (CN) individuals and patients with Alzheimer’s disease (AD), using axial MRI brain scans from the ADNI database. Multiple classifiers – LDA, QDA, KNN, SVM, and ANN – are tested, with the best performance achieved using 3D Zernike invariants and a QDA classifier with regularization, reaching an accuracy of ACC = 0.925. The extension from 2D central slices and their invariants to 3D invariants of full brain scans yields a notable improvement in both accuracy and Matthews correlation coefficient (MCC). The presented method is computationally efficient, demonstrating strong predictive capabilities without the high hardware demands typical of many current state-of-the-art deep learning models. Moreover, its generality and speed make it applicable to a wide range of 2D and 3D image classification problems beyond biomedical imaging.