PBMCs, clumps and debris, with a person in the loop.

CellCount is a general-purpose cell detector with viability. It finds cells in a brightfield image and calls each one live or dead by trypan blue exclusion. It does not classify cell types, so it will not decide on its own that an object is a red blood cell, a platelet or a fragment. What handles those is the cell-size gate, the sensitivity setting, and you, on the marked image.

That is a narrower claim than the pages that rank for these questions make, and it is the one we can stand behind on every sample rather than the uniform cell line a benchmark was run on.

A PBMC preparation that carries red blood cells will show them in the image, and a detector that counts cells will see them. Two things keep them out of the count. The cell-size gate excludes what is smaller than your cells, which removes most red cells and platelets from a preparation that is reasonably clean. And every detection is drawn on the image, so what the gate did not catch, you can. A preparation with a heavy red-cell load is a preparation to clean up first; no reader counts around it well.

Clumps are detected as aggregates and reported, but kept out of the cell count and the viability, so a clump neither inflates the count nor pulls the live share around. Where a live and a dead detection overlap, the more confident one survives. If a clump should have been counted as cells, or cells as a clump, that is one click on the image, and re-running at another sensitivity is another.

Debris is small, faint and irregular. The size gate removes what is too small for your cells; the Low sensitivity setting keeps only confident detections when an image is dirty; and the pooled cell-size histogram shows you where debris sits relative to your population before you set either.

PBMCs and other primary cells are small with faint edges in brightfield. Trypan blue makes the edges visible, and the High sensitivity setting keeps the fainter detections a uniform cell line never needs. On held-out frames the model had never seen, the live-or-dead call agreed with an expert annotator on 98.7% of the 2,682 cells both found; how it does on your preparation is what a few of your own frames will show.

Add a missed cell, delete a wrong one, flip live to dead, re-run at another sensitivity. Nothing is final until the person who knows the sample accepts the count, and the accepted count is dated, attributed to that person and sealed with a hash. The definitions of the size gate and PBMCs are in the glossary.