Enhancing cancer differentiation with synthetic MRI examinations via generative models: a systematic review

Dimitriadis, Avtantil; Trivizakis, Eleftherios; Papanikolaou, Nikolaos; Tsiknakis, Manolis; Marias, Kostas

doi:10.1186/s13244-022-01315-3

Insights into Imaging

Table 2 Details in generative methodology as presented in studies for brain tumors

From: Enhancing cancer differentiation with synthetic MRI examinations via generative models: a systematic review

Author	Architecture	Training dataset	Input	G. Arch	D. Arch	Loss function	Optimizer	Batch size	Output
Beers [38]	PGGAN	-	128-d Noise vector	nearest neighbor interpolation	CNN	W-distance	Adam	16	256 x 256
Han [39]	DCGAN/ WGAN	61,600 slices each sequence	Noise vector	Transposed CNN	CNN	GAN loss/ W-distance	Adam/ RMS-prop	64	64 x 64 128 x 128
Han [41]	PGGAN	5,036 tumor 3,853 non-tumor	Noise vector	nearest neighbor interpolation	CNN	W-distance & GP	Adam	16	256 x 256
Han [42]	PGGAN/ MUNIT/ SimGAN	154 patients 4,679 tumor & 3,750 non-tumor images	512-d Noise vector/ Synthetic images	nearest neighbor interpolation/ Style encoder & Decoder/ Refiner	CNN/ multi-scale CNN/ CNN	W-distance & GP/ MUNIT loss/ GAN loss with Self-regularization	Adam/ Adam/ SGD	16/ 1/ 10	256 x 256
Han [44]	CPGGAN	2,813 tumor images 5,963 bounding boxes 16,962 normal images	Noise vector & bounding boxes	nearest neighbor interpolation	CNN	W-distance & GP	Adam	4	256 x 256
Shin [46]	pix2pix	3,416 pairs of MRI sequences	Annotated masks	U-Net	PatchGAN	pix2pix loss	-	-	128 x 128 x 54
Chang [47]	AsynDGAN	170 patients 11,057 images	Tumor mask	9-block ResNet	PatchGAN	Custom	Adam	3	256 x 256
Chang [48]	AsynDGAN	170 patients 11,349 images	Tumor mask	9-block ResNet	PatchGAN	Custom	Adam	10	256 x 256
Deepak [49]	MSG-GAN	-	512-d Noise vector	Up-sampling & Convolutions	CNN	W-distance & GP	RMS-prop	20	128 x 128
Qasim [50]	Red-GAN	14,850 slices	Label masks	7 SPADE ResNet	PatchGAN	Hinge loss & Feature matching loss	Adam	-	256 x 256
Kwon [52]	3D-GAN	210 subjects	1000-d Noise vector	Resize-convolution / 3D-CNN	3D-CNN/ FC	Custom	Adam	4	64 x 64 x 64
Chen [54]	HybridGAN	600 slices	Fake patches & Constrained noise vector	Encoder–Decoder	CNN	Custom	Adam	30	256 x 256
Pesteie [55]	ICVAE	3,000	Real images	Encoder–Decoder	-	KLD	SGD	100	128 x 128
Hamghalam [56]	Enh-Seg-GAN	200k patches	32 x 32 patches	Recalibration block Encoder–Decoder	Markovian CNN	Custom	-	-	32 x 32
Qi [57]	SAG-GAN	225 patients	Real data	ResNet	PatchGAN	Custom	-	-	240 x 240
Guo [58]	SAMR	72 patients 1,080 instances	Lesion mask	Two encoders Residual blocks & Decoder	Six multi-scale Labelwise PatchGANs	Custom	Adam	8	256 x 256 x 5
Guo [60]	UCG-SAMR	72 patients 1,080 instances	Lesion mask & Atlases	Encoder–Decoder	Six multi-scale Labelwise PatchGANs	Custom	Adam	8	256 x 256 x 5
Ge [66]	Pairwise GAN	Mutation: 330/modality Wild-type: 672/modality	Labeled data & tumor mask	U-Net	Markovian CNN	Custom	Adam	-	128 x 128 x 4
Ge [67]	Pairwise GAN	Mutation: 33 Wild-type: 66/ HGG: 126 LGG: 45	Labeled & Unlabeled data tumor mask	U-Net	Markovian CNN	Custom	Adagrad	9	128 x 128 x 4
Carver [68]	GAN model	164 patients	Semantic labels & Real images	U-Net	VGG-19 & PatchGAN	Custom	-	-	256 x 256
Mok [69]	CB-GAN	220 HGG 54 LGG	Semantic labels	Convolution Residual block Transposed Convolution	CNN	Custom	Adam	-	-
Dikici [72]	cGANe	-	Noise vector	Transposed CNN	CNN	BCE	Adam	8	16 x 16 x 16
Kamli [75]	SMIG	17 patients/ 3,416 pairs	Lesion tumor volume / Normal volume	U-Net	PatchGAN	-	-	-	256 x 256 x 4
Li [77]	DC-AL GAN	-	100-d Noise vector	Transposed CNN	AlexNet	Custom	Adam	64	512 x 512

G. Arch, Generator Architecture; D. Arch, Discriminator Architecture; 128-d, 128-dimensional; PGGAN, Progressive Growing of Generative Adversarial Networks; CNN, Convolutional Neural Network; W-distance, Wasserstein distance, Adam, Adaptive Moment Estimation; DCGAN, Deep Convolutional GAN; WGAN, Wasserstein GAN; RMS-prop, Root-Mean-Squared propagation; GP, Gradient Penalty; MUNIT, Multimodal Unsupervised Image-to-Image Translation; SimGAN, Simulated and unsupervised images through adversarial training, SGD, Stochastic Gradient Descent; CPGGAN, Conditional Progressive Growing of GAN; AsynDGAN, Asynchronized Discriminator GAN; ResNet, Residual Network; SAG-GAN, Semi-supervised Attention-Guided GAN; SAMR, Synthesis of Anatomic and Molecular MRI images network; UCG-SAMR, Unsupervised Confidence-Guided SAMR; AdaGrad, Adaptive Gradient algorithm; MSM-GAN, Multi-Scale gradient GAN; RMS-prop, Root-Mean-Squared propagation; Enh-Seg-GAN, Enhancement and Segmentation GAN; VGG, Visual Geometry Group; CB-GAN, Coarse-to-fine Boundary aware GAN; cGANe constrained GAN ensembles; SMIG, Synthetic Medical Image Generator; DC-AL GAN, Deep Convolutional AlexNet GAN; FC, Fully Connected; ICVAE, Independent Conditional Variational Auto-Encoder; BCE, Binary Cross-Entropy; KLD, Kullback–Leibler Divergence, symbol ’-’ represents that the corresponding information was not provided in the publication

Back to article page