Real-time anomaly detection inside a live endoscopy
We built an embedded computer-vision system that flags and classifies anomalies in an endoscopic video feed in real time - during the procedure, not in a review queue afterward.
Most anomaly detection in medical imaging happens after the fact - a specialist reviews recorded footage later, which means anything time-sensitive gets caught after the moment it mattered most. In a live procedure, the person who can act on a finding is in the room right now.
The goal was to move detection into the procedure itself - an embedded system running during the exam, not a model scoring a video file hours afterward.
From labeled frames to a live-procedure assist
Four stages take the system from training data to real-time detection inside an actual procedure.
Labeled training data Annotated frames
Training data was built from annotated endoscopic frames covering multiple classes - polyps present and absent, with distinct types labeled separately.
Model training Classification
A classification and segmentation model was trained to tell those classes apart reliably, learning to distinguish each anomaly type from normal tissue.
Real-time inference Nvidia embedded
The trained model runs on embedded Nvidia hardware, processing the live gastrointestinal video feed frame by frame during the procedure itself.
Detection & integration In-procedure
Anomalies are detected and segmented in real time - flagged and classified as the feed plays, integrated directly into the procedure rather than a separate review step.
Built for the moment it matters, not after it
What it delivers
Engineering behind it
Need real-time intelligence inside a live workflow?
We build embedded computer-vision systems that run during the moment that matters, not in a review queue after it.
Book a discovery callAnonymized case study. Client details are confidential; described by domain and outcome.