Hi Taha,
Which measure did you use for bias field comparison?
If you use the correlation coefficient you can obtain a wrong estimation
since it seems to be amplitud dependent (really donīt know why)
If you use a scale invariant measure like the coeficient of variation of
the ratio of the aplied and estimated bias field you will obtain a more
logic output.
hope this helps
Jose Manjon
S. Taha AHI wrote:
>Dear SPM team,
>
>Although there are a number of similar studies in literature, I have
>downloaded synthetic T1-weighted MR volumes from Brainweb to investigate the
>accuracy of SPM2-bias field estimation. The volumes had different Gaussian
>noise but the same level of intensity non-uniformity.
>
>After masking non-cortex tissues, I applied the following 2 lines of Matlab
>code on every volume (volumes that have different level of noise but same
>amount of bias):
>======
>EstimatedBiasField = spm_bias_estimate(maskedVOLUME, [1024 1 35]);
>correctedVOLUME = spm_bias_apply(maskedVOLUME, EstimatedBiasField);
>======
>
>Then, I compared "correctedVOLUME.dat" with the ground truth bias field
>supplied by Brainweb. The aforementioned Gaussian noise levels were 0%, 1%,
>3%, 5% and 7%, and interestingly the lowest accuracy (far lowest) is
>obtained at zero Gaussian noise.
>
>I know that applying default parameters for the "spm_bias_estimate" may be
>one of the causes, but do you have another guess for this outcome ?
>
>
>Best Regards,
>- Taha
>
>
>
|