Analylizing the Correlation Between Features & Shape Changes

[Steven A]

I have one more question, once we have designed a SSM, how would you go about correlating the eigenvectors/values from said shape model with a certain feature, such as age, bmi, height, weight, etc so that you can create a correlated bare that demonstrates the changes with strong associations to that specific feature? Or, a bare such as the components from an SSM that would allow you to slide it to demonstrate the typical impacts of such a feature on a model?

[Marcel Luethi]

Hi Steven

One possibility is to use a (Bayesian) linear regression, where you set up a regression, which has the model coefficients as predictors and the feature as an outcome variable.

Another interesting possibility is the use of copulas, as described in this paper:

Egger, Bernhard, et al. "Patient-specific conditional joint models of shape, image features and clinical indicators." International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2019.

Best,

Marcel

On Thu, Jun 17, 2021 at 11:33 PM Steven A <steve...@gmail.com> wrote:

I have one more question, once we have designed a SSM, how would you go about correlating the eigenvectors/values from said shape model with a certain feature, such as age, bmi, height, weight, etc so that you can create a correlated bare that demonstrates the changes with strong associations to that specific feature? Or, a bare such as the components from an SSM that would allow you to slide it to demonstrate the typical impacts of such a feature on a model?

--
You received this message because you are subscribed to the Google Groups "scalismo" group.
To unsubscribe from this group and stop receiving emails from it, send an email to scalismo+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo/be1f9437-feba-4524-a020-8a4e86ba4603n%40googlegroups.com.

[Steven A]

Would you say that we could also use the mean model created from the SSM, then using random sampling across that mode, and each model used to discern the mean shape (that has a set feature value such as age defined), compute the distance between those models, and create a matrix where each sample has a listing of distances for each point away from the mean, and then perform a typical regression, classifier, etc? I understand that each distance would not be an orthogonal feature, but maybe you could use a statistical approach to reduce the dimensionality or a manifold approach?

[Marcel Luethi]

Dear Steven

I am no statistician and it is difficult for me to judge the validity of the procedure you suggest. Though I believe that the simplest possible analysis you could do is the linear regression and that it makes sense to start with something simple. For a linear regression, you only need to treat the model-coefficients as predictors and the attribute as the outcome variable and run a standard regression analysis. Maybe it already solves your problem. If not you can always use a more complicated problem.

Best regards,

Marcel

To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo/d9b4b957-61e3-4963-b088-71b8785d85c6n%40googlegroups.com.