Quality assurance · Production

Genentech develops breakthrough labeling process for medical imagery ML

The problem

Genentech's clinical development team needed advanced quality workflows and secure infrastructure to label medical imagery training data at scale, but relying solely on trained medical experts for annotation was costly and time-consuming, limiting their ability to deploy life-saving AI algorithms.

Workflow diagram · grounded in source
1
AI initiative triggers labeling need
trigger
“Early clinical development team launched AI initiatives in 2019 and needed advanced quality workflows for training data & a highly secure infrastructure in place immediately to enable a faster go-to-market motion”
2
Expert-guided labeler training
human_review
“having domain experts train teams of labelers on medical imagery annotation tasks”
3
Labeler annotation creation
output
“regular labelers creating annotations”
4
Expert quality sampling and review
validation
“which are then sampled and reviewed for quality by experts”
5
Labeled training data production
output
“a much faster and less expensive way to create training data and get their life-saving algorithms into production”
Reported outcome

By having domain experts train labelers instead of performing all annotation themselves, Genentech is now producing labeled data that is 10x cheaper, 5x faster, and with better quality, enabling scaled training data operations and faster algorithm deployment.

Reported metrics
Labeled data cost10x cheaper
Labeled data production speed5x faster
Labeled data qualitybetter quality
Reported stack
Labelboxconvolutional neural networks
Source
https://labelbox.com/customers/genentech-customer-story/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

By having domain experts train labelers instead of performing all annotation themselves, Genentech is now producing labeled data that is 10x cheaper, 5x faster, and with better quality, enabling scaled training data o…

What tools did this team use?

Labelbox, convolutional neural networks.

What results were reported?

Labeled data cost: 10x cheaper; Labeled data production speed: 5x faster; Labeled data quality: better quality (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

AI initiative triggers labeling need → Expert-guided labeler training → Labeler annotation creation → Expert quality sampling and review → Labeled training data production.