Towards a perturbation-based explanation for medical AI as differentiable programs