I think FID will work. It tries to match the feature distribution to be the same as the real images. So it measures how well you generate different pathological images. FID does not need to have the concept of class.