Suggestions and recommendation for improvement.

[Mark Batesole]

Hello all, 

I'm currently undergoing research using the statismo command line interfaces. We're interested in determining if it is possible to predict tooth root shape from only the crown data by fitting a posterior model. 

Our Setup: 

We've collected 75 extracted lower first bicuspids from various individuals that fit certain inclusion criteria such as: no dental fillings, no broken or missing pieces, no calculus, etc. These teeth were scanned using optical methods, and then landmarked with 22 landmarks. Using those landmarks we translated and rotated each specimen in an iterative way until all specimen were aligned to the first. These data formed the bases for our shape model. 

The test subjects consist of 25 conebeam DICOM radiographs that have an associated optical intra-oral scan of the teeth. Unfortunately trying to get a high resolution segmentation of a tooth crown is difficult using only the conebeam data. Therefore, we aligned the optical scans into the volume using landmarking and mesh fitting and then segmented the roots from the cone beam, and the crown from the optical scan. These aligned meshes were then saved as a single mesh which became our test subjet. (segmented_root_crown.png). This was repeated 25 times for a total of 25 subjects to test. 

Here are the steps we have taken using statismo: 

     1) Create GP model using the first sample as a reference: 

         statismo-build-gp-model -t shape -k gaussian -p 5 -s .1 -n 250 -r 0.vtk bicuspid_gp.h5

     2) Create datalist.txt which contains the filenames and locations of all of the aligned meshes. 

     3) Now Create a PCA model: 

         statismo-build-shape-model -p GPA -l datalist.txt -o pcamodel.h5

     4) Then extend the model using a Gaussian Process

         statismo-build-gp-model -m pcamodel.h5 -k gaussian -s 1 -p .5 -n 80 -o augmented_pcamodel.h5

     5) Fit a posterior model to a test sample that wasn’t part of the original 75 that created the pcamodel and use landmarks for only the crown. 

         statiso-posterior -i augmented_pcamodel.h5 -o sample_posterior.h5 -f sample_points.csv -m mean_points.csv -v 0.01

     6) Pull the mean from the posterior model and compare that to the test subject to which it was fit. 

         statismo-sample -i sample_posterior.h5 -m -o posterior_fit.vtk

And here is the result for the first test sample: 

(x.png, x2.png, y.png, y2.png, z.png, z2.png, iso.png)

I was hoping for a better result. Can you see something glaringly wrong with the way I'm using statismo? Would changing certain parameters result in a better fit? 

I appreciate any feedback you can give me. 

Sincerely, 

Mark Batesole, DDS, MS

[Marcel Luethi]

Hi Mark,

Your pipeline looks good and the parameters you chose seem reasonable. It is, however, difficult to judge without seeing the actual models. Did you visualize the models you get using the viewer (https://github.com/statismo/statismo/wiki/Statismo%20Viewer). When you draw random samples from the (augmented) PCA model, all the instances should look like reasonable teeth, and you should see sufficient variability to explain the training data.

What I also noticed is that you are not doing a surface fitting, but only fit to the landmarks. If you do not have any landmark information at the root of the tooth, it is not surprising that it does not lead to better results.

Did you try out the command statismo-fit-surface (https://github.com/statismo/statismo/wiki/cli-statismo-fit-surface)?

Best regards,

Marcel

--
You received this message because you are subscribed to the Google Groups "statismo" group.
To unsubscribe from this group and stop receiving emails from it, send an email to statismo-users+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.