Semisupervised Probabilistic Clustering of Brain MR Images Including Prior Clinical Information

Abstract

Accurate morphologic clustering of subjects and detection of population specific differences in brain MR images, due to e.g. neurological diseases, is of great interest in medical image analysis. In previous work, we proposed a probabilistic framework for unsupervised image clustering that allows exposing cluster specific morphological differences in each image. In this paper, we extend this framework to also accommodate semisupervised clustering approaches which provides the possibility of including prior knowledge about cluster memberships, group-level morphological differences and clinical prior knowledge. The method is validated on three different data sets and a comparative study between the supervised, semisupervised and unsupervised methods is performed. We show that the use of a limited amount of prior knowledge about cluster memberships can contribute to a better clustering performance in certain applications, while on the other hand the semisupervised clustering is quite robust to incorrect prior clustering knowledge.