Segmentation of Infant Brain Using Nonnegative Matrix Factorization

This study develops an atlas-based automated framework for segmenting infants’ brains from magnetic resonance imaging (MRI). For the accurate segmentation of different structures infant’s brain at isointense age (6–12 months), our integrates features diffusion tensor (DTI) (e.g., fractional anisotropy (FA)). A (DT) image and its region map are considered samples a Markov–Gibbs random field (MGRF) that jointly models visual appearance, shape, spatial homogeneity goal structure. The appearance is modeled with empirical distribution probability DTI features, fused by their nonnegative matrix factorization (NMF) allocation to data clusters. Projecting initial high-dimensional feature space onto low-dimensional significant NMF allows better separation structure background. cluster centers in latter determined training stage K-means clustering. In order adapt large infant inhomogeneities segment images more accurately, descriptors both first-order second-order taken into account space. Additionally, MGRF model used describe based on voxel intensities pairwise dependencies. An adaptive shape prior spatially variant constructed set co-aligned images, forming atlas database. Moreover, described uniform 3D labels. vivo experiments nine datasets showed promising results terms accuracy, which was computed using three metrics: 95-percentile modified Hausdorff distance (MHD), Dice similarity coefficient (DSC), absolute volume difference (AVD). Both quantitative assessments confirm integrating proposed NMF-fused intensity prior, DTI.