Hi, I worked on this issue (scratches) in the past, then I gave up due to lack of time.
The “leakage” is implicitly meant as “effect of the emulsion on the IR channel”, as you also figured out. It’s not a “leakage” in the sense of the electronics involved as you seem to have undestood 
When “leak” (so to say) recognition is performed, the resulting image is PERFECT, I remember doing some tests. However, it required image per image tuning.
What I never got to find out is:
- which colour channel(s) is/are best for the identification and substraction of the scene from the IR image, so that the process can be automated
- how to properly perform the substraction once the best channel(s) is/are found. As I pointed out in the other thread Difference between two regression techniques different apparently correct approaches bring different correlation parameters, except for a simple test image, only for which they match as they should!
Regarding thresholding: mentioned by @chris : it cannot work, because besides scratches and dust (relatively sharp), there is something else in the IR channel… drop-shaped or smeared-shaped residues of developing/fixing solution! They cause the histogram to be much broader and with a much less certain threshold values. It won’t work automatically at all, and trust me that even manually you’ll have a hard time choosing the value.