Hello I want to create a probabilistic graphical model for bones classification from PET/CT scan with uncertainity quantification, but I am not sure is it good idea and how to start .
So the general plan how I imagine it.
- each bone will have the class label
- there will be some variables that will be examined by separate algorithm
a) continuous parametrs describing various curvatures and size of given bone, also quantities derived from primary random variables will be analyzed like gaussian curvature from principal curvatures.
b) continuous parameter describing relative position of given bone to the neighbouring bones - this will be conditionally independent from bones located father away given nearby bones - so for example relative location of the skull may be modelled as conditionally independent from position of the pelvic bones given position of top cervical vertebra
I am able to give some prior probability distributions that will give most probable range of relative direction of one bone to the other.
Also i will specify as prior most probable values of curvatures AND some relative curvatures that will depend on the primary curvatures - hence no independence here.
In order to improve chance of convergence in learning i can restrict the possible values of some features and give the maximum and minimum possible value for each for given classes
I can model further structure using some or, and probabilistic gates that will help to group first the bones on the basis of simple rules
For example that if it has roughly similar geodesic length in both principal axes it can not be a long bone - so i can subdivide so the bones into groups like long bones, short bones, flat bones… And then subcliasify into concrete bone, reducing number of possibilities.
Thanks for help, and I welcome any ideas and critique