MIKAIA® AI Author: Create New AI and Batch-analyze an Entire Dataset

This video walks you through the entire Digital Pathology batch analysis workflow from beginning to end:

1. Load your scans into MIKAIA^®.
2. Create and train a new AI model with the AI Authort.
3. Analyze your entire dataset with your newly trained AI.
4. Review the results and exported result files.

The AI Author App is used here only as an example. A batch analysis can be conducted just the same with any of the other various apps available in MIKAIA^®, such as the cell detection apps for IHC, H&E, or immunofluorescence.

To keep it simple, in this tutorial video we show how to train a new AI model for analyzing a set of H&E stained whole-slide-images from The Cancer Genome Atlas (TCGA). We will create here a single-class AI that can recognize connective tissue. It will assign all other tissue that looks too different from the training class into the “Unsure” class. Of course, you can train your AI to detect multiple tissue classes.

The AI Author’s underlying AI technology uses a method called “Few Shot Learning”, which means to train an AI to learn something new with only few shots, i.e., few training annotations. The advantage: You do not need to draw hundreds of training annotations. And the training is blazingly fast. Still, to obtain a robust model, it is recommended to train it on multiple slides. Here we show how to do that.

Digital Pathology Batch Analysis

Once you are happy with the AI model’s performance, you can then use it to analyze new slides. In the video, we simply select all scans in the dataset and then kick-off the batch analysis by clicking the analyze “Batch” button. A job for each slide is created and added to the batch processing queue. You can lean back and let MIKAIA^® do the job. Make sure to enable “Export results” and select a target folder. MIKAIA^® will create a set of output files:

• For each slide:
  • shortcut to original scan
  • low-resolution (ca 2000x2000px) proxy images with and without burnt-in markup (for quick review)
  • spreadsheet (*.csv) with results from only that slide
  • markup file in MIKAIA’s *.ano file format (can be opened and exported to various other markup formats such as CSV, XML (Leica Aperio format) or GeoJson (loadable by QuPath)
• Once for the entire batch
  • accumulated spreadsheet (*.csv) with results from all slides. Can be opened in Microsoft Excel, or loaded into R, Python, or any other statistics tools.
  • config file that contains the used configuration parameters (for documentation and repeatibility)
  • batch analysis file (in MIKAIA’s *.micbat format) for later repeating the analysis, if required
  • in case the AI Author was used: the AI model (in MIKAIA’s *.ai format).

AI Author App in detail

See this app note: MIKAIA AI Authoring App.