There is no backlash in either belt or screw if you keep it tensioned and move it only in one direction as scanner does. CCD head is running on linear bearings on two polished quide bars and at least my Canoscan has a reduction gearbox (just a set of visible nylon gears) between stepper motor and belt so the minimun distance in one step can be as small as the designer wants.Can you imagine the jitter involved with moving a sensor strip by 0.0002"? Then there's "skew," whereby one side of the sensor strip moves a few ten-thousandths of an inch further than the other side of the sensor strip. I'd guess such a device would use a lead screw, rather than belt drive, but lead screws suffer from backlash, causing positioning to be uneven.
That being said, 20MP sensor in my m4/3 camera has square pixels in matrix setup so any image taken with that gives a very precise dpi number assuming that the real dimension of object is known. In 35mm film case it's: 36*24mm
When I fill the whole 20MP m4/3 picture with 35mm film image area (3:2 ratio) and crop it out I get 5043 * 3362 pixels which is also 17MP.
This equals 3600dpi scan resolution and as already mentioned, in my case the film grain is already becoming visible so as far as resolution is concerned, this is as good as it gets untill I find Topaz AI version that is optimized for film grain instead of digital noise. Even when I compare these new 3600dpi m4/3 images with "4800dpi scans" I took from the same negative with CanoScan, the one with more details is definitely MFT camera image.
Other people might get somewhat different results because back in the analog film era I could not afford very fast and sharp lenses so I ended up shooting with ISO200 and ISO400 film which has fairly large grain.