Applying Sparse Coding on Putamen of Discriminating Premature Newborns

Abstract

Many children born preterm exhibit frontal executive dysfunction, behavioral problems including attentional deficit hyperactivity disorder and attention related learning disabilities. Anomalies in regional specificity of cortico-striato-thalamo-cortical circuits may underlie deficits in these disorders. Here, we applied a recently developed multivariate tensor-based morphometry (mTBM) method to extract features from the surface anatomy of the striatum (putamen and globus pallidus) between 17 preterm and 19 term-born neonates scanned at term-equivalent age. For such surface-based features, the feature dimension is usually much larger than the number of subjects. We used dictionary learning and sparse coding to effectively reduce the feature dimensions. With the new features, an Adaboost classifier was employed for binary group classification. The new framework outperformed several standard imaging measures in classification, and achieved 86% accuracy, 83% sensitivity and 89% specificity. The new approach combines the efficiency of sparse coding with the sensitivity of surface mTBM, and boosts classification performance.