Hello,

I have been using HAMMER to register two skull-stripped, T1 MR images of the human brain. The source and target images have been affinely registered using FLIRT. I used FAST to segment them and changed the labels to those specified in the manual. I created the model-label image using the labelimage function. While the registration works, I would like to improve the accuracy of registration. I changed the various parameters (-i, -m, -R) but they do not seem to make a difference. The algorithm seems to converge to the same deformation field.

I calculated some standard distance metrics for registration (using ITK) for the various trials and their values remain the same (to the last decimal). I have included a few examples here:
MSE = Mean Squared Error
MI = Mutual Information
CC = Correlation Coefficient
MR = Mean Reciprocal
GD = Gradient Difference


Trail 1: Default Values: -i = 50; -R = 0.05; -m = 0.06

MSE = 140.899
MI = 1.2618
CC = 0.975279
MR = 1.07684e+07
GD = 3.57189e+07

Trial 2: -i = 50; -R = 0.07; -m = 0.15

MSE = 140.899
MI = 1.2618
CC = 0.975279
MR = 1.07684e+07
GD = 3.57189e+07


Trial 3: -i = 300; -R = 0.07; -m = 0.2

MSE = 140.899
MI = 1.2618
CC = 0.975279
MR = 1.07684e+07
GD = 3.57189e+07

Trial 4: -i = 1000; -R = 0.07; -m = 0.2

MSE = 140.899
MI = 1.2618
CC = 0.975279
MR = 1.07684e+07
GD = 3.57189e+07

Trial 5: -i = 1000; -R = 0.09; -m = 0.04

MSE = 140.899
MI = 1.2618
CC = 0.975279
MR = 1.07684e+07
GD = 3.57189e+07

Trial 6: -i = 1000; -R = 0.05; -m = 0.02

MSE = 140.899
MI = 1.2618
MR = 1.07684e+07
GD = 3.57189e+07

What should I do to improve or change the registration? Why does the registration not show any difference with changes in parameters?

Thank you.

Vidya Rajagopalan

BioSystems Imaging Lab
Virginia Tech