Training data, the raw material

An AI is only as good as its training data — where it comes from, how it is labelled and how bias slips in.

Training data is the raw material of every AI model. For breast imaging, that means mammograms, ultrasounds or slides, each labelled — ideally with biopsy-confirmed outcomes. Labels matter as much as images: bad labels teach bad lessons.

Data composition determines behaviour. If most training images come from one country, one machine vendor or one age group, the model learns that world. Diverse, representative, well-curated data — and honest reporting of what it contains — keep models fair and general.

What good training data requires

Verified, clinically meaningful labels
Diversity of populations, equipment and settings
Enough volume for the task's complexity
Clear documentation of provenance and limits

Model quality is capped by data quality. You cannot train a fair model on an unfair dataset.

In brief

Labels matter as much as the images themselves.

Data composition determines model behaviour.

Bias enters through the training data.

Keep learning

Bias and fairness in AIAI Hub Data quality and ethicsAI Hub How models are evaluatedAI Hub