Marmoset brain segmentation from deconvolved magnetic resonance images and estimated label maps

FIGURE S1 Hierarchical Bayesian model for the parameters' and hyperparameters' priors, where the deblurred image is modeled by a mixture of GGDs, the hidden label field follows a Potts MRF and the mean follows a Gaussian distribution. The parameters appearing in the boxes are fixed in advance

FIGURE S2 Showing the native axial slice with the manual GM and WM segmentation compared to the deblurred images and their respective segmentations with the noise variance being pre-estimated from the background signals, pre-estimated by a Rician noise estimator and modelled, respectively

FIGURE S3 The effect of varying the value of the granularity coefficient on the labels map produced by the JMCMC algorithm and that would be further used as prior maps for SPM segmentation

FIGURE S4 SSIM index and ISNR computed between the deconvolved image and the ground truth for different MCMC iterations and the elapsed time (in secs) for each set of iterations to be completed

FIGURE S5 Effect of varying the number of iterations on the estimated labels map

FIGURE S6 Showing simulated GM, WM and CSF regions with MoG distributions that: (A) perfectly fits a GGD distribution, (B) have a good GGD fit and (C) have a bad GGD fit

FIGURE S7 The simulated, corrupted and recovered images along with their estimated labels maps for the three cases where the regions were created with MoG distributions that: (A) perfectly fits a GGD distribution, (B) have a good GGD fit and (C) have a bad GGD fit

FIGURE S8 Showing GM, WM and CSF segmentation of the images simulated with a perfect GGD fit, a good GGD fit and a bad GGD fit compared to the true labels

FIGURE S9 An axial slice of the experimental results done on an additional real MR volume showing the segmentation results of GM, WM and CSF with segTemp and segLabels

TABLE S1 Parameters of the Gaussian distributions forming the MoG of each region

TABLE S2 SSIM index and DICE coefficients for GM, WM and CSF segmentation of the three simulated images compared to the true labels