This paper evaluates the performance of ten different methods for classifying patients with Alzheimer's disease (AD) from structural MRI data using the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The methods are categorized into three groups: voxel-based, cortical thickness-based, and hippocampus-based. Three classification experiments were conducted: distinguishing between cognitively normal (CN) controls and AD patients (AD vs. CN), identifying prodromal AD (MCI converters) from CN (CN vs. MCI converters), and predicting conversion from MCI non-converters to AD (MCInc vs. MCI converters). The results show that whole-brain methods ( voxel-based or cortical thickness-based) achieved high accuracies (up to 81% sensitivity and 95% specificity) for AD vs. CN classification. However, sensitivity was significantly lower for prodromal AD detection. No classifier showed significantly better results than chance for conversion prediction. The use of DARTEL registration improved the performance of six out of 20 classification experiments compared to SPM5 unified segmentation. Feature selection did not improve performance but increased computation times. The study highlights the importance of considering different preprocessing steps and method combinations to enhance classification accuracy.This paper evaluates the performance of ten different methods for classifying patients with Alzheimer's disease (AD) from structural MRI data using the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The methods are categorized into three groups: voxel-based, cortical thickness-based, and hippocampus-based. Three classification experiments were conducted: distinguishing between cognitively normal (CN) controls and AD patients (AD vs. CN), identifying prodromal AD (MCI converters) from CN (CN vs. MCI converters), and predicting conversion from MCI non-converters to AD (MCInc vs. MCI converters). The results show that whole-brain methods ( voxel-based or cortical thickness-based) achieved high accuracies (up to 81% sensitivity and 95% specificity) for AD vs. CN classification. However, sensitivity was significantly lower for prodromal AD detection. No classifier showed significantly better results than chance for conversion prediction. The use of DARTEL registration improved the performance of six out of 20 classification experiments compared to SPM5 unified segmentation. Feature selection did not improve performance but increased computation times. The study highlights the importance of considering different preprocessing steps and method combinations to enhance classification accuracy.