Configuring your model
Last updated
Last updated
Once you have a model that works well, you might want to add and remove different metrics for morphological characteristics, intensity, model score, or others.
You can change the metrics computed per object, predict fewer or greater objects, or create parent-child hierarchies to answer questions like "How many of class X in class Y?". Find each section below.
To customize model behavior, head over to AI Projects and click on your project. In the bottom section under Active AI models, choose the version you'd like to customize. Then follow the instructions for what you want to customize.
For object classes, every prediction comes with a model score, which is a decimal ranging from 0 to 1. To reduce the number of bad predictions, we set a threshold at 0.05, which filters out predictions with a model score less than 0.05.
To change the threshold used for cutoff during analysis time, go to the Train/test predictions tab. Simply drag the slider, and observe the changes that you see. Then, to lock in that threshold for analysis, select Update analysis threshold.
Each model comes with default metrics for each object, such as X-Y position, object area, channel intensities, and more. You can add or remove metrics by going to the Metrics tab. Check on or off metrics, and hit Save selection to lock in your settings.
X Position - Pixel X position of the centroid of the object (left to right). This is in absolute pixels of the entire image.
Y Position - Pixel Y position of the centroid of the object (top to bottom). This is in absolute pixels of the entire image.
Area - Number of pixels that the object covers.
Perimeter - Count of pixels over the outer boundary of the object.
Image origin - Filename of the image the object originated from.
Average Intensity (channel N) - Average intensity in channel N over the object area. If there are N channels in the image, there will be N channel features.
Solidity - A measurement of the overall concavity of an object. Ratio of pixels in the region to pixels of the convex hull image.
Eccentricity - A measure of how elliptical an object is. A perfect circle has an eccentricity of 0.
Major axis - The length of the major axis of the ellipse that has the same normalized second central moments as the region. This is generally a good proxy for the longest width of the object.
Minor axis - The length of the minor axis of the ellipse that has the same normalized second central moments as the region. This is generally a good proxy for the shortest width of the object.
Curved length - A good metric for the length of a long curved object (like lung cells, worms, etc). Obtained using the binary skeletonization of the mask, this is not a good fit for mostly circular objects.
Model score - The score, generally from 0 to 1, that the AI model computed for this object. This is often referred to as the confidence score.
Equivalent Diameter - The diameter of a circle with the same area as the object.
Euler Number - The total number of objects minus the total number of holes in those objects in an image.
Extent - The ratio of pixels in the object to the pixels in the bounding box of the object.
Feret Diameter Maximum - The longest distance between points around an object’s convex hull contour.
Orientation - The orientation of the object.
Perimeter Crofton - The perimeter of the object approximated by the Crofton formula in 4 directions.
Area Bounding Box - The number of pixels the bounding box of the object occupies.
Area Convex - The area of the convex hull image, which is the smallest convex polygon that encloses the region.
Area Filled - The area of the region with all holes filled in.
Colocalization: Pearson - Correlation coefficient between channel pairs
Colocalization: Manders - Overlap coefficient between channel pairs
Area threshold -For a specified channel, applies brightness/contrast settings and determines the area above or below the specified threshold. Units are in pixels or percentage. This allows for quantifying the positive signal after adjusting the brightness and contrast. In the setting below, the system will quantify all pixels above the threshold set by the brightness and contrast setting. If you set it to Below, it will quantify all pixels below the threshold.
Class Hierarchies are parent-child relationships that allow metrics such as Number of Class X within Class Y, as well as others. Examples of how you might use this as parent→child:
Organoid→cell Get the number of cells per organoid
Cell→puncta Get the number of puncta within each cell
Tissue region→positive cell Count number of positive cells in a certain region
Tumor→immune cell Get metrics for immune cell infiltration
To create a new hierarchy, start from Configure class hierarchy. From there, select Create class hierarchy. Fill out the form and press OK.
Define region of interest hierarchies between size groups to enhance precise analysis and detection within specified areas. Once configured, the child class is segmented only within the parent class, while areas outside the parent class remain unsegmented. This saves time and credits during segmentation job.
Tissue region→positive cell Positive cells are only segmented in a Tissue region
Organoid→cell Cells are only segmented with organoid
This feature allows you to expand or contract a class of objects by desired pixel amounts:
Grow: Useful for expanding cells from the nucleus, associating cytoplasmic or extracellular expressions, or connecting puncta to the nearest cell.
Shrink: Helps in removing autofluorescence or edge effects near the boundaries of a cell or tissue sample.
Options include:
Hollow: Creates a ring around the object.
Filled: Transforms the object into a filled shape.
Customizations:
Naming Classes: Assign names to different object classes.
Color Changes: Modify the color of the objects as needed.
