Comparison of Automated Volume Extraction with FreeSurfer and FastSurfer for Early Alzheimer's Disease Detection with Machine Learning

Many machine learning methods used automatic segmentation pipelines as preprocessing steps for early Alzheimer's disease detection. Most of those pipelines were time-intensive. In this work, the results of the deep learning-based FastSurfer pipeline were compared to the popular FreeSurfer pipeline. Those pipelines were used to extract volumetric features from Magnetic Resonance Imaging scans. The FastSurfer execution time was substantially smaller than the FreeSurfer execution time. Random forest models were trained based on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort and validated for an independent test set of this cohort and a subset of the external Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (AIBL) data set. The experimental results showed that models trained using the FastSurfer data set achieved similar results to the FreeSurfer models. Both, the FreeSurfer and the FastSurfer models to predict the conversion of Mild Cognitive Impaired subjects reached accuracies of 63.89% for the independent ADNI test set.