Try the model
01Demo
try it
Run the actual model on an actual cell.
This is the same ResNet-50 that produced every number on the research page, exported to ONNX and executed locally with onnxruntime-web. Your image is read into a canvas, resized to 224×224, and passed to the network without leaving your device.
The weights are stored at float16, the precision the network was trained and evaluated at. That halves the download and moves the score by up to about 0.002 against the float32 export, so treat the third decimal as noise.
Before you use this: it is not a diagnostic tool. It is a student research prototype, not a medical device, and it has never been validated in a clinic. It does not detect leukemia and it does not diagnose anyone — it assigns a similarity score to a picture of a single cell that has already been cut out of a slide. It is designed to sit in front of a qualified reviewer who examines every cell regardless, and nothing it outputs should be used to make a decision about a person.
What this accepts: one isolated white blood cell, roughly centred, on a plain background, cut out of an already-segmented smear — the kind of image C-NMC is made of. A photograph of a whole slide, a field of many cells, or anything that is not a stained cell will still return a confident-looking number, and that number will be meaningless.
1 — Load the network
0%
Tick the box above to enable.
2 — Choose a cell image
Load the network first.
what the network saw
0.000 P(leukemic), against a threshold of 0.770
0.000threshold 0.7701.000
A score, not a finding. This tool ranks cells for a reviewer who reads all of them; it does not clear a cell and it does not diagnose anyone.
the research
At the 0.770 threshold it missed 243 of 1,094 leukemic cells in the test set — sensitivity 0.778. An unflagged cell has not been cleared; the reviewer still reads every one.
How the model was built and tested, every measurement behind it, and where it does not apply are set out in full on the research page.