Dear Vy,
you've got a similar problem than the one reported by Maximilien a few
days ago:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=SPM;lLZs6A;20100205173816%2B0000
so I suggest that you first download and install the most recent SPM8
updates (ReML has been robustified recently).
That said, there seems to be a problem when specifying an explicit mask
and disabling the thresholding (or setting it very low) - some voxels
seem to be picked up that perturb the covariance estimation. I would
need to be able to reproduce the problem in order to get a feeling of
what is going on - can we see that off-list?
Best regards,
Guillaume.
Vy Dinh wrote:
> Dear SPMers,
>
> We are having problems with the first level model estimation for SOME of our subjects
> when we use specmask.m to modify spm's default mask.img. When this happens, we get
> the following error message:
>
> Temporal non-sphericity (over voxels) : ...REML estimation
> ReML Block - 1
> ReML Iteration : 1 ...3.502273e+23
> ReML Iteration : 2 ...4.630762e+07
> ReML Iteration : 3 ...3.140066e+14
> ReML Iteration : 4 ...2.756596e+09
> ReML Iteration : 5 ...1.812705e+13
> ReML Iteration : 6 ...1.133403e+08
> ReML Iteration : 7 ...5.675346e+14
> ReML Iteration : 8 ...1.227583e+12
> ReML Iteration : 9 ...-9.321451e+16
> ReML Block - 2
> ReML Iteration : 1 ...5.721858e+22
> ReML Iteration : 2 ...7.813068e+06
> ReML Iteration : 3 ...3.199120e+13
> ReML Iteration : 4 ...4.122943e+09
> ReML Iteration : 5 ...1.537731e+16
> ReML Iteration : 6 ...2.973038e+06
> ReML Iteration : 7 ...1.286838e+14
> ReML Iteration : 8 ...3.459647e+09
> ReML Iteration : 9 ...1.196354e+14
> ReML Iteration : 10 ...3.561691e+10
>
> ReML Iteration : 45 ...NaN
> ReML Iteration : 46 ...NaN
> ReML Iteration : 47 ...NaN
> ReML Iteration : 48 ...NaN
> ReML Iteration : 49 ...NaN
> ReML Iteration : 50 ...NaN
> ReML Iteration : 51 ...NaN
> ReML Iteration : 52 ...NaN
> ReML Iteration : 53 ...NaN
> ReML Iteration : 54 ...NaN
> ReML Iteration : 55 ...NaN
> ReML Iteration : 56 ...NaN
> ReML Iteration : 57 ...NaN
> ReML Iteration : 58 ...NaN
> ReML Iteration : 59 ...NaN
> ReML Iteration : 60 ...NaN
> ReML Iteration : 61 ...NaN
> ReML Iteration : 62 ...NaN
> ReML Iteration : 63 ...NaN
> ReML Iteration : 64 ...NaN
> ReML Iteration : 65 ...NaN
> ReML Iteration : 66 ...NaN
> ReML Iteration : 67 ...NaN
> ReML Iteration : 68 ...NaN
> ReML Iteration : 69 ...NaN
> ReML Iteration : 70 ...NaN
> ReML Iteration : 71 ...NaN
> ReML Iteration : 72 ...NaN
> ReML Iteration : 73 ...NaN
> ReML Iteration : 74 ...NaN
> ReML Iteration : 75 ...NaN
> ReML Iteration : 76 ...NaN
> ReML Iteration : 77 ...NaN
> ReML Iteration : 78 ...NaN
> ReML Iteration : 79 ...NaN
> ReML Iteration : 80 ...NaN
> ReML Iteration : 81 ...NaN
> ReML Iteration : 82 ...NaN
> ReML Iteration : 83 ...NaN
> ReML Iteration : 84 ...NaN
> ReML Iteration : 85 ...NaN
> ReML Iteration : 86 ...NaN
> ReML Iteration : 87 ...NaN
> ReML Iteration : 88 ...NaN
> ReML Iteration : 89 ...NaN
> ReML Iteration : 90 ...NaN
> ReML Iteration : 91 ...NaN
> ReML Iteration : 92 ...NaN
> ReML Iteration : 93 ...NaN
> ReML Iteration : 94 ...NaN
> ReML Iteration : 95 ...NaN
> ReML Iteration : 96 ...NaN
> ReML Iteration : 97 ...NaN
> ReML Iteration : 98 ...NaN
> ReML Iteration : 99 ...NaN
> ReML Iteration : 100 ...NaN
> ReML Iteration : 101 ...NaN
> ReML Iteration : 102 ...NaN
> ReML Iteration : 103 ...NaN
> ReML Iteration : 104 ...NaN
> ReML Iteration : 105 ...NaN
> ReML Iteration : 106 ...NaN
> ReML Iteration : 107 ...NaN
> ReML Iteration : 108 ...NaN
> ReML Iteration : 109 ...NaN
> ReML Iteration : 110 ...NaN
> ReML Iteration : 111 ...NaN
> ReML Iteration : 112 ...NaN
> ReML Iteration : 113 ...NaN
> ReML Iteration : 114 ...NaN
> ReML Iteration : 115 ...NaN
> ReML Iteration : 116 ...NaN
> ReML Iteration : 117 ...NaN
> ReML Iteration : 118 ...NaN
> ReML Iteration : 119 ...NaN
> ReML Iteration : 120 ...NaN
> ReML Iteration : 121 ...NaN
> ReML Iteration : 122 ...NaN
> ReML Iteration : 123 ...NaN
> ReML Iteration : 124 ...NaN
> ReML Iteration : 125 ...NaN
> ReML Iteration : 126 ...NaN
> ReML Iteration : 127 ...NaN
> ReML Iteration : 128 ...NaN
> Failed 'Model estimation'
> Output argument "V" (and maybe others) not assigned during call to
> "/Applications/fmri_progs/spm8/spm_reml.m (spm_reml)".
> In file "/Applications/fmri_progs/spm8/spm_spm.m" (v3151), function "spm_spm" at line
> 876.
> In file "/Applications/fmri_progs/spm8/config/spm_run_fmri_est.m" (v2928), function
> "spm_run_fmri_est" at line 69.
>
> Does anyone have an idea of the problems that this error could point to or any
> suggestions on how to troubleshoot this issue? My PI suggested to look for artifacts in the
> raw data, but I couldn't find anything thus far using SPM's display and Sue Gabrieli's
> Contrast Movie function. I have ruled out many potential causes (i.e. wrong inputs,
> corrupted files, different settings). However, I am certain that the estimation failure is
> related to the masking threshold since we only get this problem when we changed the
> threshold (either through specmask.m or changing SPM's default.mask.threshold).
>
> Initially, we had successfully estimated the first level models for all the subjects when we
> did not specify an explicit mask and accepted SPM's default.mask.threshold at .8.
> However, the default mask.img excluded a significant of voxels of interest, so we
> modified the mask.img using 3 methods:
>
> 1. specmask.m
> (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind06&L=SPM&P=R233923&I=-
> 3&K=1&X=363C8156D45925E39B&Y=vy_dinh%40rush.edu&d=No+Match%3BMatch%3BMa
> tches)
> This didn't work for some subjects.
>
> 2. setting an explicit mask in spm during model specification and changing
> default.mask.threshold to -inf
> (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind00&L=SPM&P=R106819&I=-
> 3&K=1&X=58833E282F8D229F41&Y=vy_dinh%40rush.edu&d=No+Match%3BMatch%3BMat
> ches)
> This didn't work for half of the subjects we processed. Only 2 had the NaNs.
>
> 3. setting an explicit mask in spm and changing default.mask.threshold to 0.05
> I don't remember which posting discussed this, but there was a lengthy topic about
> setting the threshold at 0.05, one of the authors being Tom Nichols, who cautioned
> against changing the threshold although the resulting mask image.)
> This worked for all of the subjects, and resulted in the same mask.img as created by
> specmask.
>
> From my understanding, specmask.m modifies the SPM.mat file by setting a given
> template mask image as the explicit mask (SPM.xM.VM), changing the threshold masking
> value (SPM.xM.T) from 1s to blanks, and changing the analysis thresholds for each scan
> from 80% to -inf. With the exception of the SPM.xM.T values, it appears that specmask
> and method #2 are doing the same things. This makes me wonder about how SPM is
> interpreting .T since this can shed some light on our problem.
>
> We decided to go with specmask but we are still getting the problems with the model
> estimation. I suspect that there is something wrong with the individual's data (rather than
> the method), but I'm not sure where else to look. Any hints will be much appreciated!
>
> Thanks in advance,
>
> Vy Dinh
> Research Assistant
> Department of Neurological Sciences
> Rush University Medical Center
> Chicago, IL
>
>
> Also, note that the model estimation failed when a model had 2 sessions.
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|
