Nuclei counter

How it works

This tool uses classic image analysis, not AI, so every step is visible and adjustable:

1. Smooth the image a little, to calm pixel noise.
2. Threshold: Otsu’s method picks the brightness that best separates two populations of pixels (nuclei and background). You can override it, or click the histogram. ●Established: Replicated many times; textbook knowledge.
3. Fill holes, so nucleoli and dim centres count as part of their nucleus.
4. Split touching nuclei: every nucleus pixel gets its distance to the background. Centres of nuclei are the peaks, and a watershed grows each nucleus out from its peak until neighbours meet. ●Established: Replicated many times; textbook knowledge.
5. Measure each object: area, equivalent diameter, circularity and intensity (from the original, unsmoothed pixels).

How good is it? On the two demo fields, from the public BBBC039 benchmark (U2OS cells stained with Hoechst), the default settings find about 95% and 88% of the experts’ annotated nuclei with matching outlines (F1 scores). The second field is harder: clustered and dying cells are where simple methods struggle most. Always check the outlines by eye.

For crowded tissue or unusual shapes, deep-learning tools such as Cellpose and StarDist usually do better, at the cost of larger downloads or a Python setup. ◕Strong evidence: Several independent studies agree.

Tips

• Use the nuclear channel (DAPI or Hoechst, often blue) and an image that is not saturated.
• Enter the pixel size from your microscope to get areas in µm².
• If two nuclei are counted as one, lower Min. distance between centres; if one nucleus is cut in two, raise it.
• 16-bit TIFFs are analysed at full precision; colour images are split into channels.

Demo images: BBBC039v1 (Caicedo et al. 2019), from the Broad Bioimage Benchmark Collection (Ljosa et al. 2012), CC0. Converted from 16-bit to 8-bit for the web; the expert masks are used only to score the result.