Table 2.

Overview of VAE-based architectures for medical image augmentation, including hybrid status of architectures (if applicable), indicating the combination of VAEs and GANs used in each study.

Reference	Architecture	Hybrid Status	Dataset	Modality	3D	Eval. Metrics
Classification
[81]	ICVAE		Private	MR		Acc., Sens., Spec.
				Ultrasound		Dice, Hausdroff, …
[51]	CVAE		OpenfMRI, HCP	MR	✓	Acc., Prec., F1
			NeuroSpin, IBC			Recall
[82]	GA-VAE		ADNI, AIBL	MR	✓	Acc., Spec., Sens.
[85]	MAVENs	Hybrid (V + G)	APCXR	X-ray		FID, F1
[61]	IntroVAE	Hybrid (V + G)	Private	MR		-
[86]	DR-VAE		HCP	MR		-
[64]	VAE-GAN	Hybrid (V + G)	Private	MR		Acc., Sens., Spec.
[87]	VAE		Private	MR		Acc.
[88]	RH-VAE		OASIS	MR	✓	Acc.
Segmentation
[89]	VAE-GAN	Hybrid (V + G)	Private	Ultrasound		MMD, 1-NN, MS-SSIM
[90]	AL-VAE	Hybrid (V + G)	Private	OCT 1		MMD, MS, WD
[83]	PA-VAE	Hybrid (V + G)	Private	MR	✓	PSNR, SSIM, Dice
						NMSE, Jacc., …
Cross-modal translation
[78]	CAE-ACGAN	Hybrid (V + G)	Private	CT → MR	✓	PSNR, SSIM, MAE
[91]	3D-UDA		Private	FLAIR ↔ T1 ↔ T2	✓	SSIM, PSNR, Dice
Other
[92]	CVAE		ACDC, Private	MR	✓	-
[92]	CVAE		Private	MR	✓	Dice, Hausdorff
[93]	Slice-to-3D-VAE		HCP	MR	✓	MMD, MS-SSIM
[94]	GS-VDAE		MLSP	MR		Acc.
[80]	VAE-CGAN	Hybrid (V + G)	ACDC	MR	✓	-
[95]	MM-VAE		UK Biobank	MR	✓	MMD
[96]	DM-VAE		Private	Otoscopy		-

¹ OCT stands for “esophageal optical coherence tomography”. V = variational autoencoders, G = generative adversarial networks.