A feature selection method based on a neighborhood approach for contending with functional and anatomical variability in fMRI group analysis of cognitive states
Affiliations: [a] Centro de Investigación en Inteligencia Artificial, Universidad Veracruzana, Xalapa, Veracruz, México | [b] Departamento de Fisiología, Facultad de Medicina, Universidad Nacional Autónoma de México, México
Correspondence:
[*]
Corresponding author: Efrén Juárez-Castillo, Centro de Investigación en Inteligencia Artificial, Universidad Veracruzana, Xalapa, Veracruz, México. E-mail: juarezefren@hotmail.com.
Abstract: The study of cognitive processes performed by the human brain has greatly benefited from new technologies able to infer neuronal activity by means of noninvasive methods. This is the case of functional magnetic resonance imaging. Digital image analysis and interpretation techniques have also contributed greatly to the exploration and understanding of these brain functions. Among these techniques, the use of machine learning algorithms with the ability to automatically classify cognitive states has been particularly fruitful. In general terms, these techniques identify brain regions involved in specific cognitive processes by correlating experimental stimulation patterns with the magnitude of observed neural activity. An issue with this kind of analyses arises when comparing activation results from different subjects. This occurs due to functional and anatomical variability among individuals, even after this variability is reduced during the registration process performed on the images as part of the preprocessing. In this paper we propose a feature selection method to contend with this variability. The basic idea consists in defining the activity of a voxel (feature) as a weighted vote of the observed activity of its neighbors located at a periphery defined by an isotropic three-dimensional space. Such influence is determined by a Gaussian radial function. This approach allows comparing results among different individuals, assuming that functionally equivalent activities are not necessarily presented in the same spatial position. The results show that this spatial tolerance allows a classification accuracy of 96% (considering a threshold of ± 2 voxels, equivalent to ± 8 mm.) against the 84% obtained by a traditional feature selection method.
