Learn how Meta’s Segment Anything Model (SAM) powers Centaur Labs’ auto-annotation tools to deliver more accurate labels faster.

‍

The advancement of medical AI hinges on accurate ML models trained on trustworthy ground-truth data.  Centaur Labs’ unique crowd-labeling model follows a three-step process; we develop a set of Gold Standard ground truth labels in collaboration with customers, our expert crowd generates multiple opinions on each image in as little as 24 hours, and our algorithm scores and aggregates the top labelers based on their performance against Gold Standards. Our new auto-annotation tool accelerates the creation of these Gold Standards and minimizes human error by leveraging Meta’s open-source Segment Anything Model (SAM), enabling customers to launch projects faster than ever. Read more to learn how our SAM integration works and how auto-annotation can improve your ML model. 

‍

Introducing Centaur Labs’ auto-segmentation feature powered by SAM 

‍

Polygon segmentation is traditionally a tedious task of drawing dozens of single points around objects of interest within an image. SAM minimizes that work to a two-click action to instantly draw an accurate boundary around your objects. 

‍

‍

SAM is an open-source foundation model for auto-segmenting images with best-in-class generalization capabilities and state-of-the-art technology in image segmentation. Its ability to adapt to a wide variety of visual contexts makes it the tool for medical imaging, where diversity in data types is common. 

‍

We developed our auto-segmentation capabilities by integrating with SAM based on its proven effectiveness across a variety of medical images from fundus images to chest X-rays. Centaur Labs’ auto-segmentation feature is now available to all customers using web labeling. 

‍

‍

How to use auto-segmentation for data labeling

‍

Our auto-annotation tools are now live in our desktop labeling platform and available to Centaur Labs customers. New customers can request a demo with our team to see this tool in action. Existing customers can follow this step-by-step guide to begin auto-segmenting polygons today. 

‍

1. Navigate to a Polygon Segmentation task in an Image Project
2. Under Add Labels, click “Begin Web Labeling”
3. Select the “Auto Segmentation” lightning bolt icon to activate auto-segmentation
4. Identify an object in the image that you would like to segment
5. Draw two points on the image, “bounding” the object
6. Instantly, our AI service will generate a polygon segmentation for you
7. Repeat as many times as you like on the same image!
8. Refine points on your image as needed
9. Right-click and select Delete to get rid of a segmentation

‍

Benefits of auto-segmentation for data labeling 

‍

Traditional polygon segmentation is a tiring, painstaking process, which can increase the risk of human error creating a negative ripple effect in your immediate results and long-term ML algorithm. Our new tool combines the skills of the expert human in the loop and AI-assisted segmentation to create Gold Standard ground truth labels faster with a focus on segmentation refinement rather than generation. 

‍

Medical data labeling is a unique challenge and the stakes are very high. We understand the importance of keeping expert humans in the loop to produce accurate labels that will shape the future of medical AI and improve patient outcomes. Our tool improves the workflow and throughput of labelers while minimizing the risk of human error in a critical phase of the task setup process. 

Customer benefits of auto-annotation 

• Accelerated project timelines - Every labeling project must begin with establishing Gold Standard ground truth labels. Auto-segmentation accelerates the ground truth labeling process, enabling customers to advance their labeling projects on a shorter timeline. 
• Quickly launch tasks to our expert crowd- Centaur Labs is the only data labeling platform that leverages the wisdom of an expert crowd to source multiple opinions on every image. The sooner customers establish Gold Standards, the sooner Customers tap our expert network to label their data. 
• More accurate ML training data - More accurate Gold Standards and ground truth labels help refine your ML algorithms in the long term. Auto-segmentation helps reduce the risk of human error in the Gold Standard labeling process, thus leading to tighter quality controls of crowd-labeled data, and more accurate ML models.. 

Labeler benefits of auto-annotation 

• Improved user experience - whether you’re using an internal team to label your data or leveraging our expert crowd capabilities, auto-segmentation is faster and easier than traditional segmentation. 
• Faster annotation - Labelers can work more quickly and accurately, focusing on the refinement of AI-assisted polygon segments rather than creating segments from scratch. 
• Improved segmentation accuracy - Eliminating the tedium of each segmentation task enables labelers to focus on improving segmentations. Get the most out of your expert labelers and focus them on high-impact precision work.

‍

Sample data types improved by auto-segmentation 

‍

We know firsthand from working with a wide range of medical AI customers that there are vast amounts of data types and formats that can be leveraged to advance the field of medical AI. See how pathology and video frame segmentation can benefit from auto-segmentation.

‍

Pathology 

Pathology images can be massive, and locating and segmenting cell structures can take even the best pathologists hours per slide. Auto-segmentation enables labelers to complete tasks in a fraction of the time, especially on pathology slides containing dozens or hundreds of objects of interest.

‍

‍

Video frame segmentation for object tracking 

In video frame segmentation tasks, labelers view a series of still images taken from a video clip to identify anatomical structures or foreign bodies, E.g. surgical tools. Identifying and segmenting the same object across many frames can be tedious and time-consuming. Auto-segmentation enables labelers to repeatedly segment structures across time series data much more quickly and accurately.

‍

The use cases listed here are just a preview of how our auto-annotation tools can accelerate your labeling projects. If you don’t see your use case listed, connect with our sales team and engineers to learn how your unique use case can benefit from auto-segmentation. Book a call to learn more.