Hi Derek,
In classical model selection (e.g. AIC, BIC), multiple models are fitted, a metric is computed for each model and then the metrics are used to assess the best model.
With ARD, you do indirectly a Bayesian model comparison, but you do not have to fit all candidate models. Instead, you fit the most complex model *only* and the ARD prior penalises a-priori model complexity (i.e. a-priori no crossings are supported). If such complexity (i.e. crossing) is supported by the data, the likelihood term will dominate the ARD prior term and a crossing will be estimated.
This is done automatically, during the MCMC sampling after an initial number of iterations (burn-in), during which convergence is achieved.
To summarise, you always fit the most complex model. If you ask for two fibres, the ball and 2-stick model will be fitted. In regions where multiple fibers are supported the distribution described by merged_f2samples will be away from zero. In regions with one fibre, the ARD prior will drive the f2samples to almost zero values. For these cases, you will have some values for merged_th2samples and merged_ph2samples, but these are completely random, as (th2,ph2) are degenarate when f2->0. Probtrackx has a threshold on f2 and f3 (5% by default) and do not consider orientations that have volume fractions lower than this threshold.
Is this clearer?
Stam
On 1 Jul 2013, at 20:00, Derek Archer wrote:
> Hi Stam,
>
> Thanks for the quick response. I've read the papers you've suggested, and I have one question.
>
> ARD is used to determine if the model is a one fiber or multiple fiber, correct? After this, how is the primary fiber and secondary fiber treated differently? I'm having a bit of trouble understanding how bedpostx runs different calculations on these two groups if a voxel was probable to have crossing fibers.
>
> So in short, what's the difference in getting merged_th1samples and merged_th2samples?
>
> Thanks,
> Derek
>
>
>
|