Evaluating a fractal features method for automatic detection of Alzheimer’s Disease in brain MRI scans : A quantitative study based on the method developed by Lahmiri and Boukadoum in 2013

Abstract: The field of computer-aided diagnosis has recently made progress in the diagnosing of Alzheimer's disease (AD) from magnetic resonance images (MRI) of the brain. Lahmiri and Boukadoum (2013) have research this topic since 2011, and in 2013 they presented a system for automatic detection of AD based on machine learning classification. Their proposed system achieved a classification accuracy of 100% (2013, p. 1507) using support vector machines with quadratic kernel classifiers. The MRI scans were first translated to 1-dimensional signals, from which three features were extracted to measure the signals self-affinity. These three features were Hurst’s exponent, the total fluctuation energy of a detrended fluctuational analysis and the same analysis’ scaling exponent. The results of their study were validated using a dataset of 23 MRI scans from brains with AD and normal brains. This report makes an attempt at implementing the method proposed by Lahmiri and Boukadoum in 2013 and evaluating its accuracy on a dataset of 120 cases, out of which 60 are cases of AD and 60 are normal cases. The results were validated using both leave-one-out cross-validation and 3-fold cross-validation. A dataset of 23 cases consistent with Lahmiri and Boukadoum’s in size was considered and the larger dataset of 120 cases. The best classification accuracy for the small and large were obtained from the 3-fold cross-validation was 78,26% respectively 65,00%. The results of this study are to some extent similar to those of Lahmiri and Boukadoum’s, however this study fails to verify how their method performs on a larger dataset, as their results for a small dataset could not be reproduced in this implementation. Thus the results of this report are inconclusive in verifying the accuracy of the implemented method for a larger dataset. However this implementation of the method shows promise as the accuracy for the large dataset was fairly good when comparing to other research done in the field.