ChipQC Guardian

A quality gate for organ-on-a-chip brightfield frames. Each frame gets one of three outcomes: PASS (use it without a person looking), REACQUIRE (not a usable observation) or REVIEW (a person decides), with the evidence behind it. The model runs entirely in your browser: nothing you open here is uploaded.

Code, model and results · Technical report · Kaggle writeup

Loading…

Pick a frame

Nine real frames from two acquisition dates that the model in this page never saw during fitting.

Or open your own frame (PNG, JPEG)

At most this share of passed frames may be bad: 10% 15% 20%

Frame, as the model sees it (1792 × 1344 px).

Evidence map: blue regions pull the score towards good, red towards bad. Each of the 12,288 patches has an exact share of the score; the map averages to it.

Acquisition descriptor	Value	Limit

What you are looking at

Acquisition gate. Five physical measurements decide whether the frame is a usable observation at all.
Culture-quality score. The whole frame is read as 4 × 4 tiles by a frozen self-supervised vision transformer (DINOv2 ViT-S/14); a linear classifier on the unit-length average patch descriptor, taken after its last four blocks, gives a calibrated probability that experts would call the culture good.
Decision. A frame passes only above a threshold fitted, on other acquisition dates, to keep the share of bad frames among passed frames under the target, with a safety margin. Everything else goes to a person.

Limits

The score estimates what cell-biology experts called a good or bad culture in one laboratory's dataset. It is not a measurement of viability, barrier function or drug response, and it is not for clinical or regulatory decisions.
On a different camera the model loses about ten points of AUROC until local labelled frames are added. Treat results on your own images as exploratory.
Low scores are not reliable enough to discard a culture, so the system never rejects automatically.