Julio Ernesto Villalon Reina1, Talia Miriam Nir2, Sophia Thomopoulos2, Lauren E Salminen3, Neda M Jahanshad2, Rutger Fick4, Matteo Frigo5, Rachid Deriche5, and Paul M Thompson2
1USC Imaging Genetics Center, University of Southern California, Los Angeles, CA, United States, 2USC Stevens Neuroimaging and Informatics Institute, University of Southern California, Marina del Rey, CA, United States, 3USC Imaging Genetics Center, University of Southern California, Marina del Rey, CA, United States, 4Therapanacea, Paris, France, 5Athena Project Team, Inria Sophia-Antipolis Méditerranée, Université Côte d'Azur, Nice, France
1USC Imaging Genetics Center, University of Southern California, Los Angeles, CA, United States, 2USC Stevens Neuroimaging and Informatics Institute, University of Southern California, Marina del Rey, CA, United States, 3USC Imaging Genetics Center, University of Southern California, Marina del Rey, CA, United States, 4Therapanacea, Paris, France, 5Athena Project Team, Inria Sophia-Antipolis Méditerranée, Université Côte d'Azur, Nice, France
We found that DTI diffusivity and MC-SMT measures showed the highest prediction accuracy, but differential anatomical distributions of classifying voxels. MC-SMT may offer greater sensitivity and specificity to MCI as MC-SMT resulted in the highest recall and fewest classifying voxels.
Figure 1. Anatomical distribution of logistic regression voxel-wise weights (i.e., odds) that contribute to correct classification, for the four best classifying dMRI measures.
Table 1. Classification accuracy and MCI recall for all tested measures. Recall values are multiples of 10% as there were 10 MCI individuals in the test group. As the training and test group sizes increase, these measures may differ more from one another.
