The data have been blurred (slightly) when interpolated to the +acpc grid. This blurring cannot be un-done by back transforming to +orig. If all you really have from this unfortunate patient is this +acpc dataset, then since those 3D images are only rotated (not scaled) from the no-longer-available +orig files, you should just DICOM-ize what you have. No extra transformation will add anything. The only thing you would get by back-transforming to +orig would be the slice orientations would be the same as acquired before, rather than with the head tilted up so the AC-PC line is horizontal-ish. And unless you know the slice thickness as acquired, you won't get the actual slice positions from before, just a set of slices parallel to the original slices. And the back warping will introduce a slight amount of further blur.
I hope this isn't for some legal purpose. If that's the case, trying to get the original DICOM from the PACS or some backup device, if at all possible, would be worth the effort.