[Camino-users] snr estimation strategies

[Ian Malone]

Hi,

I used the average of the registered B0 volumes as the reference for SNR 
in both comparisons, using exp(ln(S0)) for the wdt-SNR only shifts it by 
less than +/-1 in most areas (larger differences around the edges of the 
brain and ventricles), the percentile SNR values from 10-90% are within 
0.1 or less with those using the average of the B0.

Thanks for your suggestions, good to have some reassurance and a 
possible reason for the difference.

Ian

On 29/08/13 20:21, Philip A Cook wrote:
> It's a good question. When you look at the wdt-SNR, is that using the estimated ln(S0) signal output by wdtfit? I would expect that to have a different variance because it's an average value, estimated from the weighted LLS fit to all the DWI data. The difference in GM might come from the fact that on average, the DWIs will have a higher weight compared to the b0 measurements, whereas in WM many of the measurements will have low signal and hence low weight.
>
> I think your intuition is correct, the variance of the actual measurements is the quantity that you'd need for deconvolution.
>
>
> On Aug 29, 2013, at 2:00 PM, Ian Malone wrote:
>
>> Hi,
>>
>> I'm looking at ways of estimating SNR from diffusion data and currently getting different answers from the weighted linear fit (from modelfit -noisemap) and estimating directly from the standard deviation of our multiple B0 (9 volumes). In both cases applying a FSL/BET brainmask and 3x3x3 median filtering to average B0 and sd estimates before taking the ratio for SNR. The following numbers are for one scan, but having looked at others it seems to be representative in relative terms.
>> Looking at percentile derived-SNR across the brain the SNR is consistently higher using the weighted linear fit derived values, at 10% B0-estimated SNR is 13.5, wdt-SNR is 19.9, they actually diverge slightly, at 50% B0-SNR 23.3, wdt-SNR 34.2, at 90% B0-SNR 33.7, wdt-SNR 52.3 (the ratio hangs around 1.5).
>> Masking both SNR volumes at the 10th percentile shows similar areas are excluded by both: brainstem and pons and around the striatum. Directly comparing SNR and sd maps shows that most of the difference seems to come from grey matter areas where the sd estimated from B0 volumes is higher than that from modelfit's estimation, resulting in a lower SNR (nearly a factor of two in these areas). White matter areas are closer in value.
>> My major reason for wanting to estimate SNR is as an input to spherical deconvolution for tractography. Therefore it would seem the white matter value is probably more relevant than the overall value, though this does still leave the choice of B0- or wdt-SNR. I feel the B0 sd value is the more secure choice, since it's a more direct estimate, but is there any reason the wdt-SNR might be more appropriate for this application?
>>
>> Thanks for reading,
>> Ian
>>
>> _______________________________________________
>> Camino-users mailing list
>> Camino-users at www.nitrc.org
>> http://www.nitrc.org/mailman/listinfo/camino-users
>