. 2021 Mar 1;16(3):e0247839. doi: 10.1371/journal.pone.0247839

Table 1. Literature usage of CXR in a single database for each class.

Paper	Number of distinct databases			Model/application characteristics Assumptions/Limitations
Paper	No-finding	COVID-19	Pneumonia	Model/application characteristics Assumptions/Limitations
Sedik et al. (2020) [37]	+	+	-	• Focused on DA methods.
				• Small dataset (56 images).
				• Used database not specified.
Haque and Abdelgawad (2020) [38]	1	1	-	• Separate two train/test sets:
				Dataset1: two classes, one dataset in each class
				Dataset2: two classes, one dataset in each class
Minaee et al. (2020) [40]	1	1	-	• Use of transfer learning.
Minaee et al. (2020) [40]	1	1	-	• Evaluation of unbalanced test dataset using accuracy.
Narin, Kaya, and Pamuk (2020) [41]	1	1	-	• Use of transfer learning.
Narin, Kaya, and Pamuk (2020) [41]	1	1	-	• Test dataset extremely small (10 images for both classes).
Wang et al. (2020) [42]	1	2	-	• Use of transfer learning.
Wang et al. (2020) [42]	1	2	-	• Test in images already used on train.
Hemdan, Shouman and Karar (2020) [43]	1	1 (same as ‘no-finding’	-	• Use of transfer learning.
Hemdan, Shouman and Karar (2020) [43]	1	1 (same as ‘no-finding’	-	• Test dataset extremely small (10 images for both classes)
Horry et al. (2020) [44]	1	1	1 (same as ‘no-finding’)	• Use of transfer learning.
Horry et al. (2020) [44]	1	1	1 (same as ‘no-finding’)	• Only binary classification
Abbas, Abdelsamea, and Gaber (2020) [45]	1	1	1 (same as COVID-19)	• Use of transfer learning.
				• Evaluation of unbalanced test dataset using accuracy.
				Use of DA before train/test split.
Khan, Shah and Bhat (2020) [46]	1	1	1 (same as COVID-19)	• Use of transfer learning.
Ozturk et al. (2020) [47]	1	1	1 (same as ‘no-finding’)	• Use of transfer learning.
Loey, Smarandache and Khalifa (2020) [48]	2^*	1	2^* (same as ‘no-finding’	• Use of transfer learning.
Loey, Smarandache and Khalifa (2020) [48]	2^*	1	2^* (same as ‘no-finding’	• Do not explicitly define which databases are used in each class.
Apostolopoulos and Mpesiana (2020) [49]	2^*	2^*	2^*	• Use of transfer learning.
				• Absence of test dataset.
				• Only training performance is provided.
Wang, Lin and Wong (2020) [50]	1	4	1^** (same as ‘no-finding’)	• Use of a pretrained model on ImageNet images.
Ucar and Korkmaz (2020) [51]	1	1	1 (same as ‘no-finding’)	• Use of transfer learning.
Ucar and Korkmaz (2020) [51]	1	1	1 (same as ‘no-finding’)	• Use of DA before train/test split.
This work	4	4	4	• Dedicated CNN (does not require transfer learning).
				• Use of DA after train/test split.
				• Use of Balanced Accuracy (BA) metric for the unbalanced test dataset.
				• Analysis of the effect of image dimension.

* Authors did not clearly mention which databases were used in each class. Therefore, this number is an upper bound.

** More than 99% of this class came from a single database.

+ Not specified