TABLE 6. Sizes of the Datasets Used in the Reviewed Research.

Research study	Total samples		Training samples		Testing samples		Method for handling class imbalance
	ALL	COVID-19	ALL	COVID-19	ALL	COVID-19
CovidAID [90]	3,969	106	3,516	80	424	19	Fixed ratio of classes in each batch
[72]	3,550	864
[30]	109,895	225
Deep-COVID [28]	5,071	71	2,031	31	3,040	40
[73]	5,941	68	4,753		1,188
[74]	37,220	314	3,883		11,706	314
COVID-CAPS [31]	94,638	315		284		31	Modification of the loss function
[26]	1,531	100	764	50	767	50
COVID-Net [60]	13,975	358					Batch re-balancing
DeTraC [36]	196	11	138		58
[75]	15,085	180	3,783	149	11,302	31	Fixed ratio of classes in each training batch: # samples = 633
[56]	930	310					Equal # samples for each class = 310
CheXNet [59]	2,339	187	2,159	127	180	60	Data augmentation
COVID-DA [62]	11,663	318	10,718	258	945	60	Focal loss
[34]	800	240	800	200	160	40	Equal # samples for each class = 200
CoroNet [53]	18,529	99	16,576	89	1,953	10	Class-weighted entropy loss function
[76]	621	207	366	125	255	82	Equal # samples for each class = 2017
[57]	8,588	62					SMOTE
[41]	327	129
[8]		127		125
[5]	9,672	161	7,254		2,418		Data augmentation
[92]	610	324	500	250	110	74	Equal # samples for each class
[46]	50	25	50	20	50	5	Equal # samples for each class
[35]	316	158	0.6		0.2
[44]	455	135	204	102	251	33	Equal # samples for each class = 102
[79]	5,949	1,536		153		153	Data augmentation
[91]	1,493	284		228		56
[58]	6,320	464					Class-weighted entropy loss function
[80]	224,316
[81]	16,130			10		10
[82]	1,141						New conditional loss, learned with joint balance optimization and cost-sensitive learning
[42]	860	260	580	180	140	40