Table 3.

Summary of the top five combinations of pretext tasks, patch sizes, and sampling methods for each dataset with the corresponding p-value for each combination, and sorted by p-value in ascending order. The bolded pretext task, patch size, and sampling method were chosen as the best combination of the three design choices.

MRI
Rank	Pretext Task	Patch Size	Sampling Method	p -Value
1	Context Restoration	64 × 64	Random	1.64 × ${10}^{-18}$
2	Context Restoration	8 × 8	Random	1.89 × ${10}^{-18}$
3	Context Restoration	32 × 32	Poisson-Disc	1.05 × ${10}^{-17}$
4	Context Restoration	8 × 8	Poisson-Disc	4.03 × ${10}^{-17}$
5	Context Restoration	32 × 32	Random	9.38 × ${10}^{-16}$
CT
Rank	Pretext Task	Patch Size	Sampling Method	p -Value
1	Context Restoration	32 × 32	Poisson-Disc	6.29 × ${10}^{-17}$
2	Context Restoration	64 × 64	Poisson-Disc	1.88 × ${10}^{-9}$
3	Context Restoration	32 × 32	Random	4.12 × ${10}^{-7}$
4	Context Prediction	64 × 64	Poisson-Disc	0.1
5	Context Prediction	64 × 64	Random	0.66