Nevertheless, these methods typically use the whole set of voxels to compute a prediction and, so, it is difficult to threshold the weights and interpret them in terms of their role importance in the patient condition. In addition, when exploited with linear kernels, SVM provide weights for each voxel enabling the visualisation of brain patterns linked to the diagnosis ( Vemuri et al., 2008 Zhang et al., 2011). The success of this method in this domain is due to its competitive performance when the number of features is large in comparison with the number of samples. One of the most commonly used ML methods in neuroimaging is Support Vector Machines (SVM) ( Hearst et al., 1998). Feature selection presents in general the benefit of keeping the results interpretable, unlike feature extraction methods such as partial least squares ( Wold et al., 1984 Geladi and Kowalski, 1986) or principal component analysis ( Jolliffe, 1986). Due to high dimensionality issues, it is often necessary to use feature reduction methods before the learning process in order to improve performance ( Chu et al., 2012 Segovia et al., 2012 Mwangi et al., 2014). Researchers exploit these information with machine learning algorithms to achieve the best possible predictive performance or sometimes to learn more about the brain areas involved in the studied disease. While structural magnetic resonance imaging (sMRI) modality is helpful to detect brain atrophy from MCI to AD ( Jack et al., 1999 Killiany et al., 2000), functional MRI and fluorodeoxyglucose positron-emission tomography (FDG-PET) highlight function and metabolism alterations of the brain ( Chételat et al., 2003 Rombouts et al., 2005). Machine learning (ML) methods have been increasingly used over the years in neuroimaging in general and in particular also for the design of prognosis systems for Alzheimer's disease (see Rathore et al., 2017 for a review of classification frameworks designed for AD and its prodromal stages). Many studies have focused on this prodromal stage of Alzheimer's disease ( Petersen et al., 1999, 2001). Before a definitive AD diagnosis has been established clinically with neuropsychological tests, individuals go through a stage of “mild cognitive impairment” (MCI) during which predicting the outcome, stabilisation or worsening of the cognitive deficit, is difficult. Nervertheless, it still remains a challenge to predict if one individual will develop the disease before brain damages and irreversible symptoms have already appeared. As current clinical trials testing amyloid-modifying therapies in demented individuals failed to show any effect, it is believed that interventions must start before the onset of clinical symptoms ( Sperling et al., 2014). Much research has been undertaken in order to find treatments to delay the onset of the disease or slow down its progress ( Hardy and Selkoe, 2002 Roberson and Mucke, 2006). Then, they are applied on our own dataset of FDG-PET scans to identify the brain regions involved in the prognosis of Alzheimer's disease.Īlzheimer's disease is currently the neurodegenerative disease the most often encountered in aged population and, as the world's population ages, the prevalence of the disease is expected to increase ( Brookmeyer et al., 2007). The good behaviour of these methods is first assessed on artificial datasets. We then adapt several permutation schemes to turn group importance scores into more interpretable statistical scores that allow to determine the truly relevant groups in the importance rankings. Assuming a prior division of the voxels into non overlapping groups (defined by an atlas), we propose several procedures to derive group importances from individual voxel importances derived from Random Forests models. In particular, we investigate the benefit of group-based, instead of voxel-based, analyses in the context of Random Forests. In this paper, we focus on the ability of these methods to provide interpretable information about the brain regions that are the most informative about the disease or condition of interest. Machine learning approaches have been increasingly used in the neuroimaging field for the design of computer-aided diagnosis systems. 3GIGA-CRC in vivo Imaging, University of Liège, Liège, Belgium.2GIGA-CRC in silico Medicine, University of Liège, Liège, Belgium.1Department of Computer Science and Electrical Engineering, Montefiore Institute, University of Liège, Liège, Belgium.Marie Wehenkel 1,2 * Antonio Sutera 1 Christine Bastin 3 Pierre Geurts 1 † Christophe Phillips 1,2 †
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |