Multi-Camera Splitter

A splitter sits between the tube lens and the cameras, divides the beam with dichroics, and sends each spectral band to its own sensor. Every channel is exposed at the same instant rather than one after another.

For a fixed slide, that is worth very little. Nothing is moving; acquiring channels sequentially gives you the same picture and costs only time.

It is worth a great deal in three situations, and it is worth naming them because they are the only ones that justify the price. Anything moving — a beating heart, a motile cell, blood flow — decorrelates between sequential frames, so a ratio computed across channels acquired 200 ms apart is a ratio of two different moments. Anything photobleaching unevenly biases sequential ratios, because the channel you acquire second was exposed during the first. And anything where the photon budget is the limit benefits, since one excitation event can be read by several detectors at once instead of being spent on one channel and repeated.

If your samples are fixed, flat and stable, a splitter is buying you throughput and nothing else, and there are cheaper ways to buy throughput.

This is the cost, and it is the one most often discovered after purchase.

Two sensors cannot be mounted to share a pixel grid. They differ in position by tens of micrometers, in rotation by fractions of a degree, and in effective magnification by a fraction of a percent because their arms contain different glass. None of that is a defect; it is the tolerance of physical parts.

The consequence is that channels must be registered computationally, from a calibration made with multi-spectral beads, beads that fluoresce in every channel, so the same object appears in all of them and the transform between them can be solved. Without that step, a colocalization result is a statement about the mounting hardware.

The number to hold on to: a misregistration of half a pixel, at a typical sampling of 0.11 µm per pixel, is 55 nm. That is smaller than the resolution limit and therefore invisible in the image, and it is the same order as the distance a colocalization claim is asserting. An error you cannot see, at the scale of the thing you are measuring, is the most expensive kind.

It also drifts. The registration you solved in the morning is not the registration you have in the afternoon, because the instrument has warmed by a degree and aluminum moves about 23 parts per million per degree. Re-calibrate on the same cadence you would re-focus.

Each dichroic in the chain reflects one band and transmits the rest, and neither operation is free. Reflection typically runs 95 to 98%, transmission 90 to 95%, and a four-way splitter needs three dichroics to make four arms.

The first channel is peeled off at the first surface and sees one interaction. The last channel passes through all three. At 92% per transmission, that is 0.92³ = 78%: the far arm starts better than a fifth down before its camera has done anything at all. Add the emission filter in front of each sensor and the spread between best and worst arm widens further.

So a splitter's loss quoted as one number is an average, and no single channel sees it. A five-channel splitter losing 10% on average, with the same loss at each of its four dichroics, passes about 96% on its first channel and about 86% on its last two. Compare that instrument against another and the answer depends on which channel you compared: at the resolution element, a few percent either way is the whole margin between two good systems. Measure each channel's transmission, and quote comparisons per channel. The Sampling Comparator on the sample area per pixel page lets you step through the channels.

Which has a design consequence worth acting on: put your dimmest marker in the shortest optical path, not wherever it lands by wavelength convention. The arm order is usually fixed by the dichroic set, but the assignment of dye to channel is yours, and getting it backwards costs you the signal you were least able to spare.

Then there is the quiet one. Each camera has its own gain, its own read noise, its own population of hot pixels, its own flat-field. A single camera with a filter wheel has one of each. A four-camera system has four, all of which drift independently, and a quantitative comparison across channels is only as good as the worst-maintained of them.

A splitter divides light the only way light can be divided without a filter wheel: by color. So it can only separate dyes that are separable by color in the first place, and it does nothing whatever about spectral overlap.

Worse, it removes your best remedy. The Spectral Path Visualizer on the dichroic mirror page makes the size of this concrete: Alexa 594 with Atto 565 alongside it runs at about 85% bleedthrough with every laser line on at once, and falls to single figures acquired sequentially, because sequential acquisition switches off the other dye's laser, and nothing else can. A splitter is simultaneous by construction. That 85% is not a setting you can change; it is the instrument you bought.

So the decision order matters, and it is the reverse of how it usually goes. Choose the panel first and let it tell you whether a splitter is admissible. Dyes 25 nm apart want a filter wheel and sequential acquisition. Dyes 100 nm apart can be split simultaneously and the splitter will pay for itself.

Buying the splitter first and choosing the panel afterwards means discovering that the fastest instrument in the building cannot cleanly acquire the panel the biology wanted.