RBC AI Live demo (opens in a new tab)

See what an AI model sees in a blood smear.

RBC AI is a research project that finds cells in microscope images, sorts them into seven types, and highlights the regions of the image that influenced each prediction. It is built for study and education.

Open the live demo (opens in a new tab) View source on GitHub (opens in a new tab)

For research and education. Not a clinical diagnostic tool.

Stylized microscope field Red blood cells, one white blood cell and several platelets, with detection boxes and a Grad-CAM style heat map over the white cell. neutrophil erythrocyte erythrocyte
Stylized illustration of a smear field. Not real model output.

Not for clinical use. RBC AI has not been clinically validated. Do not use it to diagnose, treat, or make decisions about anyone's health.

What the project does

Four parts work together, and each one can be inspected on its own in the live demo.

  • Detection

    A YOLO-nano object detector, trained on the TXL-PBC dataset, finds white cells, red cells, and platelets across a whole smear field and draws a box around each one.

  • Classification

    A transfer-learned CNN, MobileNetV3-Small by default, assigns each cell to one of seven classes. Confidence scores are temperature-scaled, and weak predictions are marked low confidence instead of being forced into a label.

  • Grad-CAM explainability

    Grad-CAM lays a heat map over the cell, hinting at which pixels pushed the classifier toward its answer. It is a clue to the model's behavior, not proof that it reasoned correctly.

  • Evaluation

    Each trained model is scored on held-out test data, with per-class precision, recall, and F1. The demo's Research / Metrics page shows the measured numbers, and this site quotes none of them.

Seven cell classes

A class counts as supported only when expert-annotated data trained it and an independent test set evaluated it. The drawings are stylized and not to scale.

  • Erythrocyte

    Biconcave disc with no nucleus and a pale center. The most numerous cell in a smear.

  • Neutrophil

    Nucleus split into three to five connected lobes, with pale, finely granular cytoplasm.

  • Lymphocyte

    Small cell with a large, round, dense nucleus and a thin rim of cytoplasm.

  • Monocyte

    Largest white cell, with a kidney-shaped or folded nucleus and grey-blue cytoplasm.

  • Eosinophil

    Usually a two-lobed nucleus and coarse orange-red granules that fill the cytoplasm.

  • Basophil

    Dark purple-blue granules, often large enough to cover the nucleus.

  • Platelet

    Small, anucleate cell fragment, often seen in clusters between red cells.

From image to inspection

The steps run in the live demo, never on this page.

  1. Upload an image

    Choose a microscope image of a smear field, or a single-cell crop, in the demo.

  2. Detect cells

    The YOLO detector locates each cell in the field and returns a bounding box.

  3. Classify each cell

    The CNN assigns a class and a temperature-scaled confidence, or marks the cell low confidence.

  4. Inspect the results

    Review per-class counts, the annotated image, Grad-CAM evidence, and measured morphology for a selected cell.

Research and limitations

The project's own rule is no fabrication: every prediction and metric in the app is meant to come from a real model run or a real measurement.

Data and models

PBC
Acevedo et al. 2020, Data in Brief, CC BY 4.0. Backbone for the white cell and platelet classes.
Raabin-WBC
Kouzehkanan et al. 2022, Scientific Reports. Adds domain diversity.
TXL-PBC
Gan, Li and Wang, arXiv:2407.13214. Whole-smear detection training and red cell crops. See the repository's dataset registry for license details.
BCCD
MIT license. Supplementary detection data.
Segmentation
Classical computer vision (Otsu thresholding with watershed-style morphology), not a deep network.

Known limitations

  • Not clinically validated. Results must not inform any medical decision.
  • Training used public datasets. Stains, scanners, and sample preparation in other labs differ, so performance on other images can be worse.
  • Red cell morphology subtypes are not supported, because no adequate public dataset was found.
  • Whole-smear detection depends on trained detector weights. When they are unavailable, the app says so and shows no placeholder detections.
  • Models run on CPU only, and none of them replace review by a qualified professional.

Research and education only

RBC AI is not a medical device and makes no diagnostic claims. It exists so students and researchers can study how an image model reads blood cells. If you have health concerns, see a qualified clinician.