{"id":386,"date":"2023-06-13T21:45:09","date_gmt":"2023-06-13T19:45:09","guid":{"rendered":"https:\/\/websites.fraunhofer.de\/smart-sensing-insights\/?p=386"},"modified":"2025-10-14T11:16:59","modified_gmt":"2025-10-14T09:16:59","slug":"mikaia-ai-authoring-app","status":"publish","type":"post","link":"https:\/\/websites.fraunhofer.de\/smart-sensing-insights\/mikaia-ai-authoring-app\/","title":{"rendered":"MIKAIA AI Authoring App"},"content":{"rendered":"\n

MIKAIA\u00ae<\/sup> University<\/a> Application Note: Interactively training an AI with the AI Authoring App — programming-free and from just a few example annotations<\/h2>\n\n\n\n

By design, researchers often have unique questions that they want to answer as quickly and efficiently as possible. The MIKAIA\u00ae<\/sup> <\/a>team will never be able to create dedicated AI Apps for all these unique research questions. Instead, we have developed the AI Author App<\/strong> that empowers users to create an AI themselves on their own data. Since the app is based on \u201cFew Shot Learning\u201d, meaning the underlying backbone-AI that was already pre-trained on millions of histology image patches can be adapted to new use cases based on only very few examples (and in fact is not actually changed), the \u201ctraining\u201d process is very fast. Simply by defining a set of distinct tissue classes and outlining prototypical areas for each class, a new AI can be trained within minutes.<\/p>\n\n\n\n

Input<\/strong><\/strong><\/td> <\/strong><\/td><\/tr>
Staining<\/td>The features produced by the pre-trained backbone
work optimally on HE-stains.
While we observed also good performance on other stains,
such as IHC, the app cannot currently
be used with multiplexed
immunofluorescence (mIF) slides.<\/td><\/tr>
Microscopy
mode<\/td>		Brightfield<\/td><\/tr>
Magnification<\/td>Typically 20x or 40x.<\/td><\/tr>
Supported
Analysis Modes<\/td>		ROI, FoV, Slide, Batch<\/td><\/tr>
Outputs<\/strong><\/td> <\/strong><\/td><\/tr>
Graphical
Outputs<\/td>		– Polygons (with holes) that outline tissue classes
– When pre-clustering is disabled, tissue outlines are rectangular
– Optionally, a rejection class (\u201cunsure\u201d) can be activated.
When activated and depending on the sensitivity (1-100)
areas that look unlike any of the training annotations
will be assigned to the rejection class. <\/td><\/tr>
Slide-level
output metrics<\/td>		– mm\u00b2 and % per tissue class
– T-SNE\/UMAP\/PCA plot of patches (helps locating outliers)
– Tissue area in mm\u00b2<\/td><\/tr>
Description<\/strong><\/td> <\/strong><\/td><\/tr>
Pre-processing<\/td>Tissue detection
[Optionally] pre-clustering of similarly looking regions.
AI analysis is then used to classify each cluster.
The average cluster side length (\u00b5m) is configurable.<\/td><\/tr>
Post-
processing \/
additional
options<\/td>		Adjacent clusters\/patches of identical tissue type can be fused<\/td><\/tr>
Technology \/
Algorithm<\/td>		The app conducts a patch-wise classification,
i.e., each patch (~50 x 50 \u00b5m) is assigned to
a tissue class. An optional pre-clustering step
ensures that predicted tissue areas align smoothly
with the image contents (no checkerboard).
This approach entails that an entire slide
can be analyzed fast, typically within less
than 5 minutes (depending on GPU),
but it also means it is not well suited
to segment fine structures such as vessels
or cells that are smaller than a single patch.  

The app is AI-based and utilizes \u201cFew Shot
Learning\u201d, more precisely ProtoNets. The
ProtoNet backbone-AI has already been
pre-trained on millions of H&E histology
patches. The training was conducted in a
special way, improving generalizability to
other use-cases.<\/td><\/tr>
Speed<\/td>Seconds per field-of view.
Typically < 5 minutes per whole-slide
(depending on the GPU).
When no CUDA-enabled GPU is
available, the AI will run on the CPU,
which is slower.<\/td><\/tr>
Use cases<\/td>Detect and measure areas of interest,
such as tumors or metastases
Create regions of interest (e.g., tumor)
for downstream analysis (e.g., Cell
detection within tumor).\u2026<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

DIY — do it yourself!<\/strong> Train your own classifier on your data in four simple steps.<\/p>\n\n\n\n

    \n
  1. Define names of tissue classes you want to distinguish<\/li>\n\n\n\n
  2. Annotate some typical regions for these classes in one or multiple slides. Since the AI has been pre-trained on histology images, this step does not require too many annotations. Then adapt the pre-trained AI on your use case by analyzing the annotated regions.<\/li>\n\n\n\n
  3. Now apply your own classifier on new unseen regions or slides. If you are not happy with the accuracy yet, go back to step 2 and annotate falsely classified regions. Iteratively, the app will guide you where to annotate.<\/li>\n\n\n\n
  4. Enter a description text, pick an image or icon, and add the trained AI as a new app to the app center. It can now be used by you or your colleagues. You can even share it with other MIKAIA\u00ae<\/sup> users.<\/li>\n<\/ol>\n\n\n\n

    Video<\/h2>\n\n\n\n
    \n