Multi-view visual saliency-based MRI classification for Alzheimer's disease diagnosis
Résumé
Visual inspection is the first step performed by clinicians during evaluation of medical images in image-based diagnosis. This behavior can be automated using computational saliency models. In this paper, we investigate the potential role of visual saliency for computer-aided diagnosis of Alzheimer's disease (AD). We propose a multi-view saliency-based framework to detect abnormalities from structural Magnitude Resonance Imaging (MRI) and classify subjects in a Multiple Kernel Learning (MKL) framework. The obtained saliency maps are able to detect relevant brain areas for early AD diagnosis. The effectiveness of the proposed approach was evaluated on structural MRI of 509 subjects from the ADNI dataset. We achieved accuracy of 88.98% (specificity of 94.4% and a sensitivity of 83.46%) and 81.31% (specificity of 84.22% and a sensitivity of 74.21%) classification and for respectively AD versus Normal Control(NC) and NC versus Mild Cognitive Impairment (MCI). For the most challenging classification task (AD versus MCI), we reached an accuracy of 79.8%, a specificity of 79.93% and a sensitivity of 64.02%.