🩸 How CNNs Learn to Classify White Blood Cells

About this simulation

A digitised blood film image is a grid of pixels. A convolutional neural network (CNN) slides small learned filters over that grid, building up feature maps that respond to edges, textures and shapes — like the lobed shape of a neutrophil's nucleus or the thin cytoplasmic rim of a lymphocyte. Stacking several convolution + pooling layers lets the network combine simple features into increasingly abstract ones, until a final layer scores each possible cell type.

How it works

• The stained cell on the left represents one microscope field cropped around a single leukocyte, the CNN's input image.
• Each translucent slab is a convolutional layer; the glowing tiles are its feature maps, lighting up where a learned filter finds a matching pattern.
• A pulse travels left to right through the network for every inference pass, mimicking a forward pass through the layers.
• The four bars on the right are output "logits" turned into probabilities — the tallest bar is the network's predicted class.

Controls

• Cell sample — choose the true white blood cell type shown on the slide.
• Image noise — simulates blur, poor staining or a smeared film; higher noise lowers confidence and can flip the prediction.
• Network depth — number of convolutional layers; deeper networks build sharper, more confident feature representations.
• Show feature maps — toggles the animated activation tiles inside each layer.
• ▶ Run inference — fires a fresh forward pass. Drag to rotate, scroll to zoom.

Did you know?

NHS haematology labs already use CNN-based digital morphology systems to pre-screen blood films, flagging likely cell types so biomedical scientists can focus their time on the borderline or abnormal cells a model is least confident about.