Raman hyperspectral imaging is a powerful technique that provides both chemical and spatial information of a sample matrix being studied. The generated data are composed of three-dimensional (3D) arrays containing the spatial information across the x- and y-axis, and the spectral information in the z-axis. Unfolding procedures are commonly employed to analyze this type of data in a multivariate fashion, where the spatial dimension is reshaped and the spectral data fits into a two-dimensional (2D) structure and, thereafter, common first-order chemometric algorithms are applied to process the data. There are only a few algorithms capable of working with the full 3D array. Herein, we propose new algorithms for 3D discriminant analysis of hyperspectral images based on a three-dimensional principal component analysis linear discriminant analysis (3D-PCA-LDA) and a three-dimensional discriminant analysis quadratic discriminant analysis (3D-PCA-QDA) approach. The analysis was performed in order to discriminate simulated and real-world data, comprising benign controls and ovarian cancer samples based on Raman hyperspectral imaging, in which 3D-PCA-LDA and 3D-PCA-QDA achieved far superior performance than classical algorithms using unfolding procedures (PCA-LDA, PCA-QDA, partial lest squares discriminant analysis [PLS-DA], and support vector machines [SVM]), where the classification accuracies improved from 66% to 83% (simulated data) and from 50% to 100% (real-world dataset) after employing the 3D techniques. 3D-PCA-LDA and 3D-PCA-QDA are new approaches for discriminant analysis of hyperspectral images multisets to provide faster and superior classification performance than traditional techniques.