Automatic Segmentation of Multi-Spectral MR Brain Images Using a Neuro-Fuzzy Algorithm

This paper proposes an efficient segmentation algorithm for magnetic resonance (MR) images of the brain using a neuro-fuzzy algorithm. We apply this algorithm to various MR images, acquired from multiple MR scanners at different times, with varying slice thicknesses and fields of view. The proposed algorithm requires a priori knowledge concerning MR images of the brain. For example, MR images of the brain are symmetric, and white matter is contiguous along both sides of ventricular regions. For effective segmentation of MR images, a hybrid method incorporating both clustering and supervised classification algorithms is employed. In order to accept variations in feature values, we employ a bell-shaped fuzzy membership function and input these fuzzified values to a back-propagation (BP) network in the final segmentation step. Simulations with various MR images show that the proposed algorithm effectively segments MR images of the brain containing ambiguous boundaries.