 |
 |
Dear Mark, Thanks very much for the detailed comments! Here is what I got by running flirt with the schedule file: >> flirt -in inputfile -ref $FSLDIR/data/standard/MNI152_T1_2mm_brain.nii.gz -init xmat_T1_to_MNI152_2mm -out test1 -schedule $FSLDIR/etc/flirtsch/measurecost1.sch 0.0669054 0.000000 -0.000000 0.000000 0.000000 0.000000 0.000000 1.000000 1.000000 1.000000 0.000000 0.000000 0.000000 Final result: 0.971940 -0.035396 0.008270 -1.957650 0.033794 0.891260 0.081227 -0.548800 -0.004905 -0.131739 0.998848 5.771490 0.000000 0.000000 0.000000 1.000000 I guess the first number in the first output row is the cost value, please kindly confirm. Best, Leo On Sat, Apr 30, 2011 at 1:06 AM, Mark Jenkinson <[log in to unmask]> wrote: > Dear Leo, > > This is a fairly standard question and the answer is that > there is no good answer! The metrics you suggest in 1 > and 2 are flawed, as you pointed out. Using the same > metric that the registration uses won't tell you much, as > it is explicitly optimised by the registration method. > However, you can get cost function values from FLIRT > by using the schedule file: $FSLDIR/etc/flirtsch/measurecost1.sch > > This returns the cost for the given -init matrix (this matrix > should be the output of the registration step you've run > previously). It is difficult to interpret the cost function values > when comparing different images, as there may be differences > in the image intensities that modify the overall cost without > necessarily affecting the accuracy of the registration. > > The best solution to assess the quality of the registration is > to inspect the results by eye, or use manually placed landmarks > and quantify the displacement errors of the landmarks after > registration. There are quite a few papers dealing with the > assessment of registration quality (often for comparing methods) > which you can read more about if you are interested. > > All the best, > Mark > > > On 30 Apr 2011, at 00:24, Yiou Li wrote: > >> Dear FSLers, >> >> I want to assess the quality of the FLIRT registration of different T1 >> images to MNI152_T1 template. Could you please advise what is a good >> numeric metric for the assessment? >> >> I am thinking of the following ones: >> >> 1. Sum of mean square error between the registered and the reference >> image (this might not be a good one because I found the intensity >> scales of the two images are quite different). >> >> 2. Determinant of the transformation matrix (I know this has been used >> to evaluate local deformations, but large deformation could result in >> good registration quality, so this metric is not relevant?) >> >> 3. The final cost function value obtained by FLIRT (corrate, mutual >> information, etc.) However, flirt doesn't output this information, is >> there easy way to get this value? >> >> Any comment on this will be appreciated. >> >> Leo >> >> PS. Congrats to new royal couple ^_^ >> >