Imaris 10.1 brings native machine learning segmentation to Imaris. Previously machine learning segmentation was possible via the open source software Labkit [Labkit2022]. Now it is seamlessly possible within Imaris itself. And it’s fast!

Machine Learning Segmentation Usage Tips

The machine learning segmentation attempts to infer rules that fit your painting. The way you paint has an influence on how the classifier can learn. Here are a few suggestions:

• Paint small regions - If you paint large regions the machine learning system needs to work through large amounts of data to infer its rules and that takes longer. You will see that it is often sufficient to paint a few small regions.
• Don’t confuse the classifier - If you make mistakes in the sense of painting foreground where you want background or painting background where you want foreground, you confuse the machine learning system. This makes learning harder.
• Clarify borders (without confusion) - Train close to borders, but not too close. You want to show the classifier where you want the border to be.
• Train several rounds - It is useful to start simple (to avoid confusion), see what the classifier picks up, and then refine.
• Be prepared to stop - The classifier cannot solve every problem. If you find that adding training data doesn’t improve the prediction, then perhaps you are trying to teach the classifier something that it cannot learn.

Just Forget the Technical Details

Of course a machine learning system has some technical details that affect how it works. On a coarse level you may want to know that it is a Gradient Boosted Trees classifier [Friedman 2001]. You might want to know that the classifier receives Derivative of Gaussian features up to second order computed at a wide range of scales [Koenderink 1987], specifically 1,2,4,8,16, and 32 times the sigma specified in the user interface. You might want to know the hyperparameters of the decision trees. And more and more. But in a sense knowing all those technical details is counter to the ease of use of the workflow. The tool is so easy to use. We have put significant effort into setting up all the technical details of the system in such a way that they will just work for a large range of situations. You don’t have to care about them.

Multiple Images and Batch Processing

One of the things you may want to do is to apply the same classifier to multiple images. The trained classifier and the corresponding training data are stored as part of the Surfaces creation parameters, which can be then used to run batch processing or be opened on a new image to continue training.

Training on Multiple Images

You can store a classifier in Favorite Creation Parameters and use it for another image. When you do so you get the option to continue training on the new image while keeping the training data from the first by checking the “Keep Training Data” checkbox.

If you choose “keep training data”, add more training data for the second image and then save the classifier it will then contain its previous training data plus those from the new image.

Length Measurements on Slicers 

On OrthoSlicer and ObliqueSlicer you can do length measurements by going to the “Measure” tab.

Pseudo Flat Field Correction 

Pseudo flat field correction is available as an image processing function in Imaris to correct for systematic intensity variations across the field of view that arise from properties of the acquisition system. A typical example is a decrease of intensities toward the edges of the field of view.

The pseudo flat field correction function uses a Gaussian Filter of user-specified width to estimate a flatfield F from the original image I. It also optionally estimates a dark field D that is otherwise defaulted to 0. With these, the corrected image C can be computed using the formula C= (I−D)×m/(F−D) where m is the image averaged value of F-D.

Original

Pseudo Flat Field Corrected