- Computer Vision
- Healthcare · Radiology
- Deployed to Production
A radiology team's image backlog was growing faster than staff could review it, and routine screening was eating time that specialists needed for complex cases. They came to us for a way to triage that queue automatically, without replacing the radiologist's judgment call on any single case.
The Challenge
Every image in the backlog still needed a human decision, but not every image needed the same amount of attention. The team needed a way to separate routine, low-concern images from ones that deserved a closer look first, without introducing a tool that could quietly miss something a radiologist would have caught.
The Approach
We trained a deep learning classifier on the team's historical, de-identified image data, working closely with radiologists to validate the model against real diagnostic judgment rather than just benchmark accuracy. The model was tuned to flag anything uncertain for review instead of making a silent call, and we built the surrounding pipeline to slot into the team's existing review workflow rather than asking them to adopt a new one.
The Result
The classifier now runs on every incoming image and flags cases for review before a radiologist opens them, cutting down the time spent on routine screening so the team can focus on the cases that need it most. It's been running in production since launch, with radiologists keeping full authority over every diagnostic decision.