Improve your knowledge of the application training process with IKOSA AI.
...
Note |
---|
Important: Non-WSI formats of 2D and multichannel images are supported up to a maximum image size of 625 Megapixel megapixels (e.g. 25,000 x 25,000 Pixel). |
...
...
📍 Project preparation rules
Use data with a high diversity of feature appearances.
Larger datasets provide a better training results. However, small and medium datasets will work just fine if you prepare your images correctly.
Greater The greater complexity of your data requires more images.
Note |
---|
Important The image size doesn't significantly impact the training process, but rather the feature size does. For example, if you use images with 512x512px with features averaging at 30px in diameter, your trained app will function effectively even on very large images, as long as the features remain around 30px. |
...
Prepare your images
How to draw ROI(s) for application training with IKOSA AI?
...
📍 Image preparation rules
🏷️ Label set
Create a label set including at least 1 label (e.g. for a specific cell, tissue, or staining type)
Info |
---|
Please note If you're training an app with multiple labels, aim to balance the number of annotations across each label. The closer the number of annotations per label is to one another, the better the application's performance will be. |
✏️Labeling, annotation and regions of interests
Use fully annotated and labeled images or ROIs to train your app.
Images | ROIs | ||||
---|---|---|---|---|---|
If you use images for application training, fully annotate at least 2 images. It means that you should annotate and label ALL features.In total, you should have at least 100 annotations per label.
| If you use ROIs for application training, you need to have fully annotated at least 1 ROI per image. It means that you should annotate and label ALL features inside and on the border of ROIs. In total, you should have at least 100 annotations per label.
| ||||
A mix of images and ROIs | |||||
If you annotate some images entirely while using ROIs on others, in step 6, ensure to select the option to train your application on ROIs. This way, if an image doesn’t have any ROIs, the system will use the whole image. |
...
Features that are present , but not annotated or labeled introduce counterfactual information into the trained application and impair the results (red arrows on the screenshot).
...
If some ROIs overlap and share the same annotations, these annotations will be counted for each ROI separately. Thus, in the training results, you will have a higher number of annotations in total, which originates from this overlap counted several times.
Moreover, the training will run through such 'multiplied' “multiplied” annotations more often than the other ones, effectively giving them a higher importance in the training process.
...
👁️ Background recognition
The background of an image is an area that DOESN’T CONTAIN the features required for the app training. You add them to help the app to differentiate between background and features to recognize, especially when they are very similar.
Background images/ROIs can be added at step 6 of the training process.
Info |
---|
FAQ Is it an obligation to have background images or background ROIs? No! The trained app will autonomously learn how the background differs from the features. Even if you do not provide separate background images or background ROIs, the app will obtain information on what the background looks like based on the spaces in between annotations. I don’t provide any background images or background ROIs. What is the sufficient amount of background area within an image or an ROI? Whether you use ROI(s) or not, the greater the amount of non-annotated background area available for application training, the better the resulting model will be. What if I don’t have images only with background? You can dedicate some images that have background and present features:
|
...
Note |
---|
Important:Ensure that both background images or background ROIs contain only background, as any unintentionally included unlabeled features introduce misleading information into the trained application. |
...
Random dataset split | Manual dataset split | ||||
---|---|---|---|---|---|
You select all images to be used in app development, and IKOSA randomly and automatically divides them into training (80% of the images) and validation datasets (20% of the images).
| Select at least 1 training and 1 validation image.
|
...
If you use images with specifically designated ROIs:
Select the '“Performing the training on ROIs'ROIs” option in the 'Improve “Improve background recognition' recognition” section.
Choose your images.
...
Info |
---|
Experience difficulties interpreting your outcomes? Check this article |
...
For a quick review of Go to the Training Evaluation section to quickly review your trained app or a comparison of results performance or compare it with another trained app one from a different training iteration for the same use case head to the Training Evaluation section.
...
Are you satisfied with the training results? | |
---|---|
Yes | No |
...