Sanity check on the classes
A cell called Świecące that measures dim, or Nieświecące that measures bright, is either a misclick or a cell whose oval or background is wrong. Neither is visible in a list of two thousand rows, so they are counted and shown on the results and export screens. Nothing is reclassified — the cells are surfaced and you decide.
Measured against the 377 cells you classified by hand: your Świecące run from 84 upward, your Nieświecące from 18 to 128 with a median of 96. So a limit of 100 flags 3 of your 125 Świecące, but 54 of your 133 Nieświecące — two in five of them, all correctly classified. At 120 it is 10 and 4. Set them where you want the check to bite.
Detection parameters
These defaults were validated against your manual measurements to within about one percent. Changing them means re-analysing, and the values are stamped into every export.
Otsu is never allowed below this, so an empty field stays empty.
Smaller objects are debris or fragments.
Larger objects are usually clumps.
How far apart seeds must be before two heads are split.
A pixel must be this far from any signal to count as background.
Rejects ragged outlines — debris and fragments.
Rejects slivers — tails and fragments.
Gaussian blur applied before thresholding.
Intensity above which a pixel counts as signal when finding background.
Performance
1 is recommended for this machine.
Review
Language
Class names stay Polish either way — they are the labels in your sheet.
Cached results
reading…
Nothing uploads
All processing runs in this browser. Images, results and review decisions stay on this machine — which is why a folder has to be re-opened after a reload, and why this works on samples that could not legally leave the building.