Table 2.

The OOD performance of training from scratch versus the pre-trained models (vanilla, SSL, and SWSL).

Pre-training	Weights	Training scenario	Hospital 1	Hospital 2	Hospital 3	Hospital 4	Hospital 5	Average
F	Random	S1	93.01	89.22	84.95	91.06	81.09	87.9 ± 4.22
F	Random	S2	92.72	90.28	82.01	90	80.71	87.1 ± 4.73
T	Vanilla	S1	98.75	96.03	94.42	96.65	90.54	95.3 ± 2.69
T	Vanilla	S2	98.62	93.6	97.06	97.19	91.67	95.6 ± 2.52
T	SSL	S1	98.52	96.92	94.8	97.46	96.61	96.9 ± 1.19
T	SSL	S1	99.18	94.98	95.09	97.79	97.21	96.8 ± 1.59
T	SWSL	S2	99.08	96.52	94.97	98.12	83.93	94.5 ± 5.37
T	SWSL	S2	99.31	96.19	97.44	98.09	89.71	96.1 ± 3.3
		Average	97.4 ± 1.95	94.2 ± 2.05	92.6 ± 4.01	95.8 ± 2.29	88.9 ± 4.48

Each column represents the OOD top-1 accuracy on the hold-out set.