AUGMENTATION.md

Augmentations

class Augmentation:
    def _augment(self, sample: Sample) -> Sample:
        raise NotImplementedError("Please Implement this method")

    def _set_metadata(self, sample: Sample) -> Sample:
        return sample

    def __call__(self, sample: Sample) -> Sample:
        sample = init_augmentation(sample)
        sample = self._augment(sample)
        sample = self._set_metadata(sample)
        return sample

Base types of augmentations

Interface	Description	Metadata Changes	Code
Augmentation	Base Augmentation class from which every Augmentation inherits	None	here
OODAugmentation	Augmentation that transforms ID images into OOD images e.g. artifacts	OOD Flag, Augmentation severity if provided	here
INAugmentation	Augmentation that does not change the OOD flag but keeps the data IN Distribution e.g. cropping or resizing of images	None	here
AugmentationComposite	An Augmentation composite is a combination of multiple augmentations e.g. pick N of a list	Dependent of parameters	here

Special types of augmentations

Interface	Description	Metadata Changes	Code
SampableAugmentation	Sample Augmentation have sampable parameters which makes it possible to run an experiment with augmentations with different parameter values e.g. different scales for artifacts	Keeps Changes of base augmentation	here
ArtifactAugmentation	An Artifact Augmentation is a realization of an OODAugmentation and therefore inherits from it. This interface comes with a method that overlays an artifact on the image	like OODAugmentation	here

FrOoDos Augmentations

Name	Short description	Interface	Code
DarkSpotsAugmentation	Creates a dark spot on the image	ArtifactAugmentation, SampableAugmentation, OODAugmentation	here
FatAugmentation	Creates a fat drop on the image	ArtifactAugmentation, SampableAugmentation, OODAugmentation	here
SquamousAugmentation	Creates a tissue piece on the image	ArtifactAugmentation, SampableAugmentation, OODAugmentation	here
ThreadAugmentation	Creates a thread on the image	ArtifactAugmentation, SampableAugmentation, OODAugmentation	here
GaussianBlurAugmentation	Blures the image	SampableAugmentation, OODAugmentation	here
ZoomInAugmentation	Crops the image and then resizes it back to original size	SampableAugmentation, OODAugmentation	here
BrightnessAugmentation	Changes brightness of image	SampableAugmentation, OODAugmentation	here
InCrop	Crops image	INAugmentation	here
InResize	Resizes image	INAugmentation	here

References

[1] Schömig-Markiefka, B., Pryalukhin, A., Hulla, W., Bychkov, A., Fukuoka, J., Mad- abhushi, A., Achter, V., Nieroda, L., Büttner, R., Quaas, A., et al.: Quality control stress test for deep learning-based diagnostic model in digital pathology. Modern Pathology (2021)