in the documentation of skimage.measure.regionprops the explanation for the return values
“The length of the major/minor axis of the ellipse that has the same normalized second central moments as the region.”
The returned value resembles the equivalent radius of the respective principal value of the second central moments.
r = 2 * sqrt(I/A)
This has less to do with an ellipse, but more with an equivalent circle for this one specific principal value. If the underlying image is an ellipse the above explanation is correct, but if not, an ellipse with this minor and major axis has not the same second central moments of the region. This can be easily verified by just trying out. I at least know one issue where this was confused and it would be nice to change the documentation. I know that in Matlab there is the same wording, so maybe just the remark to radius of gyration is already enough to warn the user.