Table 4.

MC-MedGAN configurations tested

	Hyper-parameter	Model 1	Model 2
Autoencoder	Code size	64	128
	Encoder hidden size	256, 128	512, 256, 128
	Decoder hidden size	256, 128	512, 256, 128
GAN	Generator hidden layers	64, 64	128, 128, 128, 128
	Discriminator hidden size	256, 128	512, 256, 128
	# of generator/discriminator steps	2/1	3/1

For both models we used batch size of 100 samples, trained the autoencoder for 100 epochs and the GAN for 500 epochs. We applied L2-regularization on the neural network weights (weight decay) with λ=1e-3, and temperature parameter (Gumbel-Softmax trick) τ=0.66. We tested learning rates of [1e-2, 1e-3, 1e-4]