Thanks for raising it. The anisotropic pixel spacing typical of clinical imaging means that you can’t easily do 3D measurements. It’s even worse than that: your ruler in the out-of-plane direction is likely to be at least 5× bigger than the thing you are trying to measure (like trying to measure your interpupillary distance with an unmarked metre-stick).
This is not to say don’t bother trying, just, be painfully aware of the limitations to measurement that your modality has restricted you to.
BV/TV might work, but it will be very sensitive to your segmentation. You should do a sensitivity analysis to show how BV/TV varies as a function of e.g. threshold value (or whatever other segmentation steps you take).
Another approach would be to use the greyscale values, as long as they are well calibrated. I guess your signal is going from low (bone matrix) to high (marrow fat and water). All of your pixels will be partially filled with bone matrix and marrow, and the BV/TV might be able to be estimated by a mean pixel value within your ROI. One way to calibrate is to use known pixel values in the current image, e.g. a cortical bone region as ‘100% bone’ and subcutaneous or infrapatellar fat as ‘100% marrow’. You can then estimate the pixel’s filling fraction (i.e. its BV/TV) as:
f = (p - b) / (m - b)
(please check that. The idea is to place your pixel value p on a linear scale stretching from 0 - 1 between your image’s bone value b and marrow value m. It assumes a linear relationship, which may not be correct - it depends on the MR image formation physics).
This calculation can be done with ImageJ’s built-in Image math plugins.