Segmentation of Molecular Neuroimages Using Hidden Markov Random Fields in Order to Improve the Assisted Diagnosis of Neurodegenerative Diseases

In recent years, several computer-assisted diagnosis systems for neurodegenerative disorders have been presented. A large branch of these systems analyzes neuroimaging data using machine learning algorithms and are able to estimate the diagnosis of a new subject. Due to the high dimensionality of brain neuroimages compared to the number of examples available for training these systems, a dimensionality reduction strategy is recommended. In this work, we demonstrate a method to select regions of interest from molecular neuroimages, reducing, that way, their dimensionality. It is based on Hidden Markov Random Field (HMRF) and the Bhattacharyya distance. First, the average of all the images of a given modality is calculated. Subsequently, this image is segmented by a method based on HMRF. As a result, different maps are obtained according to the intensity and neighborhood of the voxels (it is more likely that a given voxel belongs to the same map than the voxels in its neighborhood). The resulting maps are then used as a mask to segment all our neuroimages. Subsequently, the data corresponding to the most discriminative maps are selected as regions of interest. The discriminative capacity is estimated by means of the Bhattacharyya distance. Finally, a Support Vector Machine classifier is used to separate controls and patients using the selected maps as features. This approach was evaluated using two sets of neuroimages, one with Alzheimer's disease patients and healthy subjects and another one with parkinsonian patients and controls. For both datasets, the proposed method achieved higher accuracy rates than classical approaches, what suggest that segmentation based on HMRF is an interesting approach to select regions of interest from molecular neuroimaging data.