Fig. 2:

Illustration of the proposed image transformations and their learning perspectives. For simplicity and clarity, we illustrate the transformation on a 2D CT slice, but our Genesis Chest CT is trained directly using 3D sub-volumes, which are transformed in a 3D manner. For ease of understanding, in (a) non-linear transformation, we have displayed an image undergoing different translating functions in Columns 2—7; in (b) local-shuffling, (c) outer-cutout, and (d) inner-cutout transformation, we have illustrated each of the processes step by step in Columns 2—6, where the first and last columns denote the original images and the final transformed images, respectively. In local-shuffling, a different window W is automatically generated and used in each step. We provide the implementation details in Sec. 2.2 and more visualizations in Fig. D.11.