Group Sparse Representation for Prediction of MCI Conversion to AD

Early Diagnose of Alzheimer’s disease (AD) is a problem which scientists are committed to solving for a long time. As the prodromal stage of AD, the mild cognitive impairment (MCI) patients have high risk of conversion to AD. Thus, for early diagnosis and possible early treatment of AD, it is important for accurate prediction of MCI conversion to AD, i.e., classification between MCI non-converter (MCI-NC) and MCI converter (MCI-C). In this paper, we propose a group discriminative sparse representation algorithm for prediction of MCI conversion to AD. Unlike the previous researches which are based on l₁-norm sparse representation classification (SRC), we focus on how to mining the group label information which will help us to do the classification more correctly and efficiently. We apply the group label restricted condition as well as the sparse condition when doing the sparse coding procedure, which makes the sparse coding coefficients discriminative. The Moreau-Yosida regularization method is utilized to help us solving this convex optimization problem. In our experiments on magnetic resonance brain images of 403 MCI patients (167 MCI-C and 236MCI-NC) from ADNI database, we demonstrate that the proposed method performs better than the traditional classification methods such as SRC with l₁-norm and group sparse representation with l₂-norm.