Brain surface extraction from PET images with deformable model: assessment using Monte Carlo simulator

Abstract

In this study, we evaluate quantitatively the performance of the DM-DSM (deformable model with dual surface minimization) method for brain surface extraction from PET images with Monte Carlo simulated data. The DM-DSM method is based on a deformable model and has been found reliable in previous tests with images of healthy volunteers acquired with C-11-Raclopride and F-18-FDG. As the evaluation of the method with real data is challenging, it could not provide precise figures describing the accuracy of the method. In addition to evaluation, we adjust parameter values for the DM-DSM method to improve its accuracy in this study. We compare the DM-DSM method to PET brain delineation based on MRI-PET registration. For this we assume either the knowledge of the precise anatomical brain volume or we extract the brain volume from the anatomical MR image. With FDG, the DM-DSM method yielded brain surfaces of high accuracy, almost as accurate as those obtained by using image registration and the knowledge of the exact anatomy. If the precise anatomical brain volume was not known, the DM- DSM method was more accurate than the image registration based method. With Raclopride, the accuracy of the DM-DSM method was slightly lower than with FDG but it was better than the one obtained using image registration and assuming the knowledge of the anatomical brain volume. When we extracted brain volume automatically from the MR image, the sagittal sinus was excluded from the brain improving the registration accuracy and leading to better quantitative results than those obtained with the DM-DSM method.