Figure 7.

The clustering results were achieved through the utilization of two MoCo v2 losses ^{( 11 )} with a VGG13 backbone, each with a distinct set of transformations, on the Nocodazole (a), Cytochalasin B (b), and Taxol (c) image treatment subsets. One loss employs color jitter, flips, rotation, affine transformation, and random cropping, while the other uses rotations, center cropping, color jitter, and flips. The clustering results demonstrate that the phenotypes of each subset are clearly separated and represented in each cluster, as evidenced by the images closest to its centroïd.