Mammography Screening AI

The program

Radisen builds AI for medical imaging. The mammography program develops software that assists radiologists in reading mammograms for breast-cancer screening — a regulated medical device, which means the model has to clear clinical validation and regulatory approval before it can be used on real patients.

That regulatory bar changes everything upstream: data provenance, label quality, and reproducibility matter as much as model accuracy. I joined this program on the operations and data side, working alongside the ML and clinical teams rather than developing the diagnostic model itself.

What I worked on

• Data pipelinesSet up and maintained pipelines to source and organize training / validation imaging data, including datasets obtained through AIHub.
• Annotation workflowSupported the annotation process — tooling and coordination so radiologist labels flowed cleanly and consistently into the dataset.
• Data QCRan training / validation data quality checks for a separate fracture-classification model, catching mislabeled and low-quality samples before they reached training.
• Cross-border commsActed as a technical communication bridge between the Korea and Taiwan teams on data and engineering matters.
• Regulatory supportHelped coordinate steps of the FDA approval process — organizing documentation and keeping communication aligned across teams.

What I took from it

This was my first close look at how a regulated medical-AI product actually ships. The lesson that stuck: in high-stakes settings, the unglamorous data and process work — where the data came from, how it was labeled, whether it can be reproduced and defended to a regulator — is often what decides whether a model can be trusted at all.