How to specify a slide scanner

Depth of field at NA 0.75 is around one micron. A glass slide bows by tens of microns across its length, the coverslip adds its own tilt, the mounting medium varies in thickness, and the section itself is neither flat nor uniformly thick. The focal surface you actually need is a warped sheet several tens of microns deep, and you have a micron of latitude anywhere on it. Nothing else in a scanner is specified this tightly.

Two strategies exist and they fail differently. A focus map samples focus at a few dozen points first, fits a surface, then scans open-loop against it — fast, deterministic, and wrong wherever tissue thickness changes faster than the fit. On-the-fly autofocus measures at every tile, using a second sensor at a deliberate defocus or a short Z sweep, and survives folds, bubbles and torn sections at the cost of time or a second camera. Most instruments end up doing both: a map for the bulk, live correction where the map is not trusted.

What this asks of the hardware is worth stating precisely, because it is usually specified wrongly. Absolute Z accuracy is irrelevant — nobody cares what the true height is. What matters is repeatability and settle time: the actuator must land within a fraction of the depth of field and stop ringing before the exposure begins, a few thousand times per slide, without drifting over the hour. That is a closed-loop piezo specification, and it is the reason open-loop drive is not an option on the Z axis even where it is perfectly adequate in XY.

Step-and-shoot under continuous illumination never raises this question, which is exactly why it ambushes people the first time they try to go faster. An sCMOS sensor with a rolling shutter starts and stops its rows in sequence, taking a fixed span to walk from the first row to the last. Fire a short flash at an arbitrary moment and it lands inside some rows' exposure windows and outside others: the frame comes back with a bright band across it, and no amount of flat-fielding removes it, because it moves.

The fix is to make an all-rows-open window exist and then put the flash inside it. That window only appears when the exposure is set longer than the row-readout span, so the last row has opened before the first one closes. Set exposure below that span and there is no instant at which the whole sensor is listening. The alternative is a global-shutter sensor, which sidesteps the geometry at some cost in noise or full-well.

Before you design a strobed continuous scan around a camera, confirm two things on its datasheet: that it publishes the row-readout span or line period, and that it accepts an external trigger with a deterministic delay. A camera that does neither can still scan — but only by stopping, which puts the settle time back into every tile.