I have no clue how to dig into this.
Check whether this is reproducible with plain imread(), no CUDA. whatever outcome, that’s a vital bit of information.
then check existing issues (CUDA issues would be on the opencv_contrib repo). if there’s nothing suitable, create a new one.