SVCell is an innovative platform for the development of image recognition analyses for a broad range of microscopy imaging applications. SVCell’s teach-by-example interfaces provide users with a simple and intuitive way to access SVision’s image based decision technologies to create novel and high performance segmentation and decision recipes. The teaching interfaces replace traditional methods of image processing, filtering, data mining and pattern recognition, and enable non-experts to quickly teach accurate analyses for their application of interest. The teaching is encoded in a SVCell recipe, which executes the complete application processing and can be scaled for high volume processing. The recipe can be updated with additional teaching at any time. Cutting edge and high performance imaging applications can be developed with low cost, minimum risk, and limited time and effort. SVCell and its recipes can be integrated with other components to deliver a complete imaging solution.

Segmentation
Segmentation is a fundamental step in image recognition applications where the objects or large structures of interest in the images are identified as distinct from the background and each other. If the segmentation is not accurate then all follow-on measurements and analyses will suffer.

SVCell’s approach to segmentation is to detect and segment individual patterns of interest, and then later combine and refine these into the final segmentation masks. Users teach the computer to recognize patterns of interest using a drawing interface. For example, the user could teach the computer to detect adherent cells in a phase image, and then teach the computer to recognize floating or round cells. The teaching produces an enhanced image, called a confidence map (one for each individual pattern of interest), where image pixels corresponding to the taught pattern of interest have a higher value (a process called confidence mapping). The confidence map can be easily thresholded to produce a binary segmentation mask, and the mask can be modified using common operations (fill holes, remove objects, logical operations, etc.). The teaching and operations are encoded into a segmentation recipe that can be used for automated processing.

Confidence mapping handles the most difficult aspect of image recognition; transforming the original image into an enhanced image that can be easily thresholded to create an accurate segmentation mask. SVCell offers two distinct methods of confidence mapping; a general purpose technology called soft matching, and a technology optimized for puncta detection called spot mapping.

One of the strengths of the SVCell segmentation approach is that it is pattern oriented and not dependent only on image contrast. As shown in the gap junction example at right, patterns of interest can be easily extracted even though they share intensity levels with other patterns in the image; something that would be difficult to do using standard image processing techniques. It is because of this strength that SVCell is heavily used for phase contrast imaging applications where image contrast is low and traditional methods are not easily applied. Please see the Segmentation teaching page for more details.

Classification
Classification is used to discriminate and count phenotypes in the imaging application. Applications include classifying live cells from dead, classifying cell types (e.g. fetal blood cells vs. maternal cells, differentiated cells vs. progenitor cells type), and classifying types of protein localization (e.g. trans-located vs. diffuse).

Classification takes place after objects of interest have been segmented and measurements made. In SVCell sophisticated classification rules can be created by teaching directly in the images; the user simply clicks on the object overlays to assign the class (e.g. live cell) label to the object. SVCell automatically creates the classification rules. The teaching and rules are saved to a decision recipe which can be used for automated processing.

SVCell decision recipes enable accurate classification of image based phenotypes which cannot be easily characterized with simple measurements or gating rules. The recipes are easy to understand, have predictable performance, and use only the measurements defined by the user.