Graph Laplacian-Based Framework for Classifying Alzheimer’s Disease Stages Using Structural MRI

Traditional approaches for Alzheimer's Disease (AD) detection often rely on manual analysis of brain Magnetic Resonance Imaging (MRI). However, accurate multiclass classification of AD stages remains challenging due to significant inter-class similarities. These methods are not only time-consuming and subject to human error but also fail to effectively capture structural relationships, such as proximity-based interactions within the brain. To address these challenges, graph-based techniques provide a powerful alternative by representing brain MRI images as networks, enabling the modeling of both spatial and spectral relationships between pixel intensities. This study proposes a novel methodology that incorporates Spectral Graph Wavelet Transform (SGWT) features with machine learning models for classifying AD into very mild, mild, moderate dementia, and non-demented stages. Both normalized and unnormalized Laplacian matrices are utilized for SGWT feature extraction, with results showing that the unnormalized Laplacian performs better due to its ability to preserve critical localized features in MRI data. The proposed framework is evaluated using boosting classifiers such as Extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost), achieving an accuracy of 98.52% with the LGBM model. These findings highlight the potential of graph-based brain MRI analysis as an efficient and accurate solution for automated multiclass AD detection.