For educators and researchers
Research & teaching
Hemra is free to use for learning and research. It is not for diagnosis or any medical decision.
At a glance
All figures are from images in the same dataset as the training data. Accuracy on images from other hospitals, stains or microscopes has not been measured.
Data
Images come from a public dataset of 17,092 normal peripheral blood cell images collected at the Hospital Clinic of Barcelona between 2015 and 2019 [1]. Smears were stained with May Grünwald-Giemsa, each cell was photographed by a CellaVision DM96 analyzer at 360 × 363 pixels, and every image was labelled by expert clinical pathologists. Donors had no infection, blood disease or cancer and were taking no medication at the time of collection.
Hemra uses a balanced subset of four cell types: eosinophil, lymphocyte, monocyte and neutrophil. The dataset's neutrophil class includes both mature segmented cells and younger band cells. Images were held out for testing and never used in training: 3,840 training and 960 test images, 240 per class.
How it works
Hemra uses a convolutional neural network, a type of model built to recognise visual patterns. It was pretrained on general images, then trained further on labelled blood cells to tell the four types apart.
When you analyze an image, it's cropped and resized, the model gives a score for each cell type, and the four scores add up to 100%. Everything runs in your browser; no image is sent to a server.
Test images were scored through the same code the public tool uses. A prediction counts as correct when the top score matches the pathologists' label.
Results
936 of 960 test images were classified correctly (97.5%).
| Cell type | Correct | Recall | Precision |
|---|---|---|---|
| Eosinophil | 236 / 240 | 98.3% | 98.7% |
| Lymphocyte | 229 / 240 | 95.4% | 99.6% |
| Monocyte | 237 / 240 | 98.8% | 93.3% |
| Neutrophil | 234 / 240 | 97.5% | 98.7% |
Recall: of the cells that really were this type, how many Hemra found. Precision: of the cells Hemra called this type, how many really were.
Where it gets confused
Rows are the pathologists' label; columns are Hemra's top answer.
| Eos | Lym | Mon | Neu | |
|---|---|---|---|---|
| Eosinophil | 236 | 0 | 2 | 2 |
| Lymphocyte | 1 | 229 | 10 | 0 |
| Monocyte | 1 | 1 | 237 | 1 |
| Neutrophil | 1 | 0 | 5 | 234 |
The most common mistake is calling a lymphocyte a monocyte (10 of 240). Both have a large nucleus and bluish cytoplasm, and large lymphocytes can look similar to monocytes [2].
Limitations
- One source. Every image comes from one hospital, one stain and one analyzer. Images from elsewhere will likely score lower.
- Healthy donors only. Hemra has never seen abnormal or cancerous cells, so it cannot detect leukemia or any other condition.
- Split by image, not by person. The public dataset doesn't say which images came from the same person, so cells from one donor may appear in both training and test sets. This can make results look better than they would be on new patients.
- Four classes, forced choice. Basophils, platelets, red cells, or a photo of anything else will still be scored as one of the four types.
- No external validation. Hemra has not been tested by an independent lab or compared with clinicians.
Use Hemra in your class
A 45-minute lesson on what image AI can and can't do in medicine.
- Introduce (10 min). What a classifier does: it compares an image to patterns it learned. Read the How it works section together and look at the four cell types.
- Explore (15 min). Students run the tool on the four samples, then on a blurry image, a red blood cell and something that isn't a cell at all, and note what happens.
- Discuss (15 min). Use the questions below and the confusion table above.
- Reflect (5 min). Each student writes one thing a tool like this should never be used for.
Discussion questions
- Why is a similarity score not the same as the chance someone has a disease?
- Hemra still gives a confident score to a photo of a cat. What does that tell you about AI models?
- Why might lymphocytes and monocytes be the pair Hemra mixes up most?
- Hemra scored 97.5% on images from one hospital. Would you expect the same score on images from your local lab? Why?
- The model only learned from healthy people. What could go wrong if someone used it on a patient with leukemia?
- Why might it matter that Hemra runs on your device instead of a server?
How to cite Hemra
Aaron, T. (2026). Hemra: white blood cell pattern research tool [Computer software]. https://hemraresarch.teddyaaron.com/
If you use the training data, also cite the dataset [1].
References
- Acevedo A, Merino A, Alférez S, Molina Á, Boldú L, Rodellar J. A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief. 2020;30:105474. doi:10.1016/j.dib.2020.105474. Data: doi:10.17632/snkd93bnjr.1, CC BY 4.0.
- Haematologica Atlas of Hematologic Cytology. Normal peripheral blood cells.