Inquiry on Density Map #415

Hi, thanks for DiffDRR and nanodrr. We've used the former heavily for chest X-ray-to-CT work.

This is a question rather than a bug report. In diffdrr/data.py:

air = torch.where(volume <= -800)
...
density[air] = volume[soft_tissue].min()

every voxel at or below −800 HU is assigned the same scalar. Normal aerated lung parenchyma spans roughly −950 to −700 HU, so on chest CTs most of the lung lands on that floor and becomes indistinguishable from the air outside the patient.

nanodrr's hu_to_mu doesn't do this: it clamps at −1000 and applies a bilinear
water/bone model, so the variation is preserved.

I assume the DiffDRR behaviour is deliberate for the registration setting the library was built around, where bone carries the signal and NCC/MI are invariant to intensity rescaling. If that's right, would a one-line note in the docstring be welcome? It isn't obvious from the outside, and DiffDRR is still what most X-ray-to-CT work installs.

Happy to send a PR with just that.