Genentech builds convolutional neural networks to diagnose illness from real patient imagery. Labelbox's platform produced clinically expert-graded signal 10x cheaper and 5x faster.

The challenge

Genentech, a biotechnology enterprise advancing medical research since 1976, builds convolutional neural networks to help diagnose illness and support medical professionals. Classic algorithms handle "perfect" data, but real patient data is full of abnormalities — a real retina might show multiple lesions and signs of illness. Training deep learning models to find and classify those cases takes meticulously labeled medical imagery, hundreds to thousands of images. And only trained medical experts can be trusted to grade it, because an inaccurate prediction can mean misdiagnosis or loss of life. Producing that volume of expert-graded signal was costly and slow.

The approach

Genentech used Labelbox to build the expert-grading infrastructure for its diagnostic models. Domain experts trained contributors on medical-imagery annotation, contributors produced the annotations, and experts sampled and reviewed them for quality — clinical judgment encoded into the signal at scale. The early clinical development team, which launched its AI initiatives in 2019, got the quality workflows and secure infrastructure it needed for a faster go-to-market.

The outcome

Genentech scaled its signal operations by letting subject matter experts teach contributors what to look for. It now produces labeled signal 10x cheaper, 5x faster, and with better quality — getting its life-saving algorithms into production sooner.

Where this goes

Diagnostic AI is only as trustworthy as the clinical judgment it's trained on. Expert-graded ground truth is what makes a medical model safe to deploy.