Table 4.

AMI score comparison in K-Means (k = 2) clusterings of models trained with two SSRL losses versus an ImageNet pre-trained encoder on Nocodazole, Cytochalasin B, and Taxol treated cell subsets. One SSRL loss uses color jitter, flips, rotation, affine transformation, and random cropping; the other, rotations, center cropping, color jitter, and flips. Careful selection of transformation sets, tailored to desired features, enhances clustering performance in self-supervised training over supervised pre-trained models, even in small datasets.

Transformations	SSRL approach	Backbone	Nocodazole	Cytochalasin B	Taxol
Weighted combination of sets	MoCo v2	VGG13	0.51	0.66	0.52
		ResNet18	0.46	0.63	0.47
	Byol	VGG13	0.51	0.64	0.54
		ResNet18	0.47	0.61	0.48
	VICReg	VGG13	0.55	0.67	0.51
		ResNet18	0.5	0.63	0.45
Pretrained models on ImageNet		VGG16	0.34	0.55	0.36
		ResNet101	0.39	0.57	0.43