Cell Counter & Image Segmentation
Free online cell counter. Upload a micrograph and count cells automatically with area, circularity and confluence. Images never leave your device.
How it works
Uploaded fluorescence images are separated into channels and segmented with a two-level hysteresis mask — a low and a high Otsu-derived threshold, chosen on two datasets by a protocol fixed before the run and then confirmed on three more that took no part in choosing it — then de-clumped with a distance transform followed by watershed segmentation, with growth capped at 3 pixels. Each resulting object is measured for area, perimeter, circularity (4πA/P²), aspect ratio and integrated density. Spatial calibration is read from TIFF resolution tags where present; when a file carries no scale, results are labelled as uncalibrated rather than silently assuming one. An opt-in deep-learning path (a small model trained from scratch on public H&E histology data, never on another tool's weights) is also available, and is the better choice specifically for H&E tissue, where the classical watershed struggles with confluent cells that have no gaps between them.
Frequently asked questions
How do you separate touching cells in a fluorescence image?
Touching or overlapping nuclei are separated with a distance transform followed by watershed segmentation. The distance transform converts the binary mask into a map where each pixel’s value is its distance to the nearest background pixel; local maxima mark object centres, and watershed then floods outward from those seeds to place boundaries where objects meet. This resolves clumps that simple thresholding merges into one object.
Why does my cell area in µm² look wrong?
Area depends on the pixel size of your image, and it scales with the square of that value. If your image file does not record a spatial scale, an assumed value must be substituted, and a 40× image read at a 20× scale reports areas roughly four times too large. Always set the objective or pixel size before trusting µm² measurements — SciKeep flags images whose scale was assumed rather than read from the file.
How accurate is the automatic cell counting?
Measured against published hand-annotated datasets, not just claimed: F1 87% on BBBC039 (evenly stained U2OS nuclei) and 71% across BBBC038, a mix of 30-plus fluorescence, brightfield and histology experiments. On H&E tissue the built-in engine is weak (F1 in the 20s), which is disclosed rather than hidden — use the opt-in deep-learning path or a dedicated tool like Cellpose for H&E work instead. Every figure and the scripts that reproduce it are on the validation page.
Method reference
Otsu N. (1979) IEEE Trans Syst Man Cybern 9(1):62-66.