Cell Detection and Counting
Turning dot annotations into a trainable segmentation target, then recovering cell positions as connected-component centroids — the standard route to automated counting.
What it is
Counting cells by hand is slow and irreproducible, which is the stated motivation behind most of this literature. The automated approach has settled into a recognisable shape.
Annotation is a dot, not an outline. A pathologist clicks the centre of each cell of interest. This is far cheaper than drawing boundaries, and for counting the boundary is not needed. The cost is that a dot is a single pixel, which is a hopeless target for a segmentation network.
The target is the dot, dilated. Each dot is grown into a small disc, producing a binary mask the network can actually learn. This is the “dilated segmentation” formulation: a detection problem re-posed as segmentation. An alternative lineage regresses a smoothed density map instead and integrates it to get a count; the dilated-mask route keeps individual objects addressable, which matters if you want positions and not just totals.
Prediction returns to points. The predicted mask is thresholded, connected components are found, and each component’s centroid — via image moments — becomes a detected cell. Counting is then just counting components.
Evaluation has two halves. Counting accuracy asks whether the total is right; localisation asks whether each detected point is near a true one, usually by matching within a distance tolerance and reporting precision, recall and F1. A method can improve one and not the other, so both belong in any report — two errors in opposite directions cancel perfectly in a count and not at all in localisation.
The characteristic failure mode is crowding. Where cells touch, the dilated discs merge, connected components fuse, and several cells collapse into one centroid. This is the specific problem that Çakı et al., Med Image Anal 2026;113:104155 targets, by making the training loss sensitive to the arrangement of cells rather than only their pixels.
Why it matters for my work
Cell counting and IHC quantification sit at the centre of the group’s digital pathology practice. The QuPath workshop teaches exactly this — hücre sayımı and İHK kantifikasyonu — and QuPath Tooling holds the annotation and downstream deep-learning repositories that would host such a pipeline. Any Ki-67 index or similar proportion is a counting task underneath, so whatever limits counting accuracy limits the index.
The annotation economics are the part worth internalising. Dots are cheap enough that a department can plausibly generate its own training data during routine work, which is not true of boundary annotation. That makes this one of the few deep-learning tasks where an in-house dataset is realistic — see KUCell for a group that did precisely that and released it.
How it connects
Persistent Homology — the tool used to quantify whether the predicted arrangement of cells matches the truth, which plain pixel metrics miss in crowded fields.
Hard Example Mining — how training is steered toward the crowded regions where this method fails.
QuPath Annotation Workflow — where dot annotations would actually be produced and exported in this group’s setup.
Interobserver Agreement — dot placement varies between annotators, and that variability is the ceiling on measurable detection accuracy.
KUCell — a released dot-annotated dataset built by a pathology department, and the closest published analogue to what Memorial could assemble.
Weakly Supervised Learning — the contrast case: cell detection is fully supervised at the object level, whereas slide-level tasks have no such per-object labels.
Cervical Cytology Screening — the same substrate applied to a screening task: per-cell detection and classification, but hunting a handful of abnormal cells among tens of thousands rather than counting a population.
Awesome-AI4DigitalPathology — where the nucleus-segmentation canon behind this page is catalogued (HoVer-Net, Cellpose, CellViT, PanNuke, MoNuSeg, CoNIC, Mesmer, Lizard, NuCLS) — filed there under cytology rather than dense prediction, so browsing by section misses it.
Open questions
- Does the group have dot-annotated data from any project, or only region annotations? Nothing in the repo records dot-level annotation.
[unverified] - Which readouts currently in use are really counting tasks underneath — Ki-67 certainly, but the Aiforia Breast endpoint is not recorded, so its dependence on detection accuracy is unknown.
[unverified] - Is there a documented distance tolerance for calling a detection correct? That threshold drives every localisation number and is rarely stated.