In older versions of Revit this was always tricky because the geometry is parameter driven and those parameter names are not always consistent between family authors.
One solution I found was to get the GeometryElement from the symbol geometry and then deduct bounding box min values from bounding box max values. In the case of symbol geometry, the bounding box should be aligned with the representation of the symbol (regardless of how instance is manipulated in space).
i.e:
For a beam X.Max - X.Min = Width, Z.Max - Z.Min = Depth.
For a column X.Max - X.Min = Width, Y.Max - Y.Min = Depth
The above is dependent on how the family was made in the first place. So logically if you have certainty of family construction i.e. knowing what parameters drive it then extraction of these is more reliable and gives better performance than the above.
Thankfully in modern versions of Revit you have the common section profiles built in parameters so if there is confidence these have been used getting such information is as easy as it should be (from consistent parameters across organisations).
