First of all, it was recommanded by ESRI (as heard) that we should NOT conduct caching on such kind of MXD anymore, which contains image services.
So, what we are currently doing is: (1.) to combine several image services with scale dependency (manually defined) in MXD, and then (2.) to publish this MXD as map services to all desktop-based end users & minority of web applications.
Anyhow, it works very well in our cases! The performance is also not too bad!