Table 1.

All benchmark settings.

Type	Benchmark	Resolution	0/1 labels	Classes	Train set	Test/val set	Per class
Chest	Classes	32 × 32	9	20	29,000	3800	1450
		32 × 32	8	15	24,000	2850	1600
		32 × 32	5	10	20,000	1900	2000
		32 × 32	5	6	13,800	1140	2300
		32 × 32	5	4	15,600	760	3900
		32 × 32	4	2	12,600	380	6300
	Samples	32 × 32	4	3	17,850	2250	5950
		32 × 32	4	3	13,500	2250	4500
		32 × 32	4	3	9000	2250	3000
		32 × 32	4	3	4500	2250	1500
		32 × 32	4	3	3000	2250	1000
		32 × 32	4	3	1500	2250	500
		32 × 32	4	3	1200	2250	400
		32 × 32	4	3	600	2250	200
	Resolution	32 × 32	14	138	117,168	4000	256−7586
		64 × 64	14	138	117,168	4000	256−7586
		128 × 128	14	138	117,168	4000	256−7586
		256 × 256	14	138	117,168	4000	256−7586
		512 × 512	14	138	117,168	4000	256−7586
Brain	Classes	32 × 32	5	10	25,000	3000	2500
		32 × 32	5	8	24,960	2400	3120
		32 × 32	5	6	25,020	1800	4170
		32 × 32	4	4	25,000	1200	6250
		32 × 32	2	2	25,000	600	12,500
	Samples	32 × 32	5	6	32,400	3000	5400
		32 × 32	5	6	27,000	3000	4500
		32 × 32	5	6	18,000	3000	3000
		32 × 32	5	6	9000	3000	1500
		32 × 32	5	6	6000	3000	1000
		32 × 32	5	6	3000	3000	500
		32 × 32	5	6	1800	3000	300
		32 × 32	5	6	600	3000	100
	Resolution	32 × 32	6	20	117,168	4000	155−85,876
		64 × 64	6	20	117,168	4000	155−85,876
		128 × 128	6	20	117,168	4000	155−85,876
		256 × 256	6	20	117,168	4000	155−85,876
		512 × 512	6	20	117,168	4000	155−85,876

Each row defines the composition of a specific benchmark setting. After GAN training, the synthetic datasets are generated by conditioning on the real label sets, resulting in equivalent data folds. Our chest radigraph data pool consists of 117,168 (44,153) training, 15,418 (5519) validation and 14,687 (5520) test samples (patients), respectively. Our brain computed tomography scan data pool consists of 173,271 (15,133) training, 22,095 (1892) validation and 20,500 (1892) test samples (patients), respectively. The 14 binary chest X-ray labels are enlarged cardiomediastinum, cardiomegaly, lung opacity, lung lesion, oedema, consolidation, pneumonia, atelectasis, pneumothorax, pleural effusion, pleural other, fracture and support device, and no finding. The six binary brain CT scan labels are epidural, subarachnoid, subdural, intraparenchymal and intraventricular haemorrhage, and no finding. 0/1 Labels: number of binary labels. Classes: number of classes. Note: the number of classes refers to the number of unique binary label combinations. If different binary labels co-occur, we can have fewer classes than 0/1 labels. Train set: number of samples in training set. Test/val set: number of samples in each the test and validation set. Per class: number of training samples per class.