An adaptation of the dino algorithm from Facebookresearch to 5 channels, easily adaptable to any number of channels (rather than just RGB).
This work was done with Cell Painting 5-channel data in mind, but could be useful for any other medical or biological imaging modalities with greater or fewer than 3 channels.
While a simple adapation, I am sure many people working with multi-channel data such as fluorescent cell images have struggled adapting many existing frameworks with 3-channel (typically RGB) implementations. For dino there are a number of augmentation steps applied to each channel in training. Most augmentation toolboxes do no work well for multi-channel (>3) images. RGB is overwhelmingly common for datasets in computer vision beyond grayscale.
The function "NaturalImageDataset" is the main contribution here. While fairly 'hard-coded', it is a practical solution to the lack of support for multi-channel augmentations. Please let me know if there are more elegant implementations.