AI Create Training Set ( IAAICreateTrainingSet )

The AI Create Training Set function is a tool for creating training data sets for ML/AI using cropping from larger images and applying data augmentation. This function supports fluorescence or bright field, color (RGB or fluorescence overlay) and grayscale images, and annotations/ROIs (implemented) or label images (coming soon). Use cases are creation of training sets of object detection, such as with YOLO architectures, or training instance segmentation with Cellpose. The image is cropped around each annotation and exported multiple times with different rotation angles and translations. Rotation and translation are performed before cropping so each crop will result in a complete image, without borders. All detection annotations that are not substantially clipped are transferred to the cropped images. Optional intensity-related data augmentation is applied after cropping, on the cropped image set. For each run, this function creates an image series, that can be further manipulated or automatically exported. Cropped images are exported and accumulated in a folder/structure based on the export format.

* Input Image A:
Image to be cropped. All overlaid images will be used. For RGB color images use the red channel with the overlaid other channels only. For annotation/ROI operation, these images must contain ROIs with ROI annotations defining one or few object classes. Use the Image Window context menu, Image Properties window or the Appearance/Set ROI Annotation function to enter class names.

* Input Image B (coming soon)
Segmented label image

* Method= Rotate large images, then crop around ROIs
The operation is performed for annotations in the “Annotations for detection” and “Annotations for background” lists. Labels are saved only for detections annotations, but not for background ones. Large images are first rotated, and then the following operations repeat “Number of rotations” times, each time with “Rotation degree” more rotation. Zero value sets optional random rotation. For each rotation, the image is cropped off centered around each annotation “Number of translations” times, using different offsets. If “Limit translations for redundant labels”=yes, no new translations are made for annotations that are already exported while cropping around nearby annotations. With “Mirror”=yes, each rotation will be performed on a mirrored-images in addition, doubling the number of crops.

* Label export format:
“YOLO yaml+txt in folders”: Cropped images are saved \images\train or \images\val (depending on if the current position in the MD Open dialog is on this list: “Validation set (MD Open positions)”) in “Image export format”. Labels are saved in the corresponding \labels\ path as a list of normalized XYWH data. The “Export path or filename” is a *.yaml file, that can exist, listing the available annotation labels, or will be created automatically. If the *.yaml file exists, repeated runs of this function will result in consistent assignment of annotation labels to classes. Call this function by running a pipeline by the Run Pipeline… on All Stage Positions command, and specify in the “Validation set (MD Open positions)” which positions will end up in the validation set. In this way the same image or augmentations of the same image do not get into both training and validations sets.

* Simulated microscopy modality:
Data augmentation allows here varying intensities and fluorescence cross bleed similarly to normal microscopic image formation. The augmentation is performed on the cropped image series before export, in a random manner.