Overview
Key Metrics & Impact
A European medical device manufacturer approached Plavno after its internal computer vision team encountered a data bottleneck.
The company had a large image collection of rigid and semi-rigid medical products — including devices such as catheters, infusion components, and moulded plastic parts — but the images had been captured in different environments and were not suitable for direct model training. Object boundaries were often affected by reflections off plastic and glass surfaces, shadows, cropping, low contrast, and partial visibility. The annotation itself was not especially complex on a single image, but repeating it consistently across thousands of files required a much stricter process.
Plavno worked with the client to formalize the segmentation rules, prepare a representative pilot set, and establish a production workflow with separate annotation and review stages. The dataset was processed in batches so that recurring errors could be identified early rather than discovered after the full delivery.
10,000 + medical product images processed through the annotation pipeline
Pixel-level masks created across varied image conditions, including reflective plastic/glass surfaces, low contrast, and partial occlusion
Automated technical validation applied to every submitted mask
Independent quality review separated from first-pass annotation
Ambiguous images routed through a documented adjudication process
Validated batches delivered in the client's required training-data format
Image-level review, correction, and approval history retained throughout the project
Client involvement focused on rules and exceptional cases rather than full-dataset manual review
