The aim of this study is to detect pathology, such as cerebrovascular disease, in brain images by assuming that the pathology is beyond the expected morphological variability of normal images. However the construction of a statistical model of the inter-subject variability over the whole high resolution image is especially challenging due to large dimensionality. For this reason, we apply image partitioning and formulate a strictly concave likelihood function estimating pathology for each local partition. We apply a distributed estimation algorithm in order to fuse the local estimates of each overlapping partition into a globally optimal estimate that satisfies consistency constraints. The likelihood function consists of a model and a data term and is formulated as a quadratic programming problem. The assessment of the method on FLAIR brain images by receiver operating characteristic (ROC) analysis demonstrates improvement in image segmentation over two-group analysis performed with SPM.