Train your own AI with the Segmentation AI Author: Ovarian Cancer Example

This MIKAIA^® University app note demonstrates how to quickly and interactively train your own AI using just a few training annotations, all in a matter of minutes.

In the app center, select the Segmentation AI Author App, click “+” to create a new trainable AI, and pick a name. Then, add the tissue classes you want the AI to recognize. In this example. we annotated “Tumor”, “Necrosis”, “Stroma”, “Background”, “Inflammation”, and “Other”. Next, create a few small training annotations. If a class has several slightly different appearances, it is best to draw multiple small annotations. The screenshot below shows the training annotations.

Now, click “train.” The Segmentation AI Author app will use its foundation model backend to extract features from the training annotations and derive a prototypical representation of each class in the high-dimensional feature space. No actual “AI training” in the technical sense is happening here, so the training phase typically only takes less than a minute.

The features can be visualized in a t-SNE or UMAP plot:

Now that the AI model has been trained, we can put it to the test and test-analyze a region in the same or a different slide:

Here is another region:

Try it out yourself: Evaluate MIKAIA^®

If you are interested in trying this or any other of the MIKAIA^® studio apps, please reach out to us at mikaia@iis.fraunhofer.de. We would love to learn about your use case, provide a demo, and afterwards share a set of voucher codes that allow you to explore the apps that are locked in MIKAIA^® lite.

Contributors & partners

• Institute of Pathology, Erlangen University Hospital, Comprehensive Cancer Center Erlangen–European Metropolitan Area of Nuremberg (CCC ER-EMN), Bavarian Cancer Research Center (BZKF), Friedrich-Alexander-Universität Erlangen-Nürnberg, Krankenhausstrasse 8–10, 91054 Erlangen, Germany
• Department of Gynecology and Obstetrics, Erlangen University Hospital, Comprehensive Cancer Center Erlangen–European Metropolitan Area of Nuremberg (CCC ER-EMN), Bavarian Cancer Research Center (BZKF), Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg, Universitaetsstrasse 21–23, 91054 Erlangen, Germany

Funding

This MIKAIA^® extension has been kindly made possible thanks to venture capital provided by the Fraunhofer Future Foundation (Fraunhofer Zukunfsstiftung). Project: “Histology AI Author”, consortium: Fraunhofer IIS & Fraunhofer MEVIS.