Table 1.

Summary of deep learning based low-dose to full-dose post-processing implementations reviewed in this work. Details from each

source are: the neural network architecture, dimensions of the input data, additional input information, tracer, anatomical region, activity and acquisition time, the dose or time reduction factor, and the evaluation metrics used to convey performance

	Network architecture	PET input dimensions	Additional input data	Tracers	Anatomy	Activity/Acq, time (MBq, min)	Dose/time reduction factor	Evaluation metrics
[73]	CNN	2D Patch	T1	18F-FDG	Brain	(203,12)	4	PSNR, nMSE
[74]	Unet	2.5D	None	18F-FDG	Brain	(370,40)	200	SSIM, PSNR, NRMSE
[75]	Residual Unet	2D	T1, T2, FLAIR	18F-FBB	Brain	(330, 20)	100	PSNR, RMSE, SSIM, QCS, rSUV, CD
[76]	Unet	3D	CT	18F-FDG	Cardiac	(300,10)	10, 100	LVEF, ESV, EDV
[77]	Modified Unet	2.5D	LAVA	18F-FDG	Whole body	Site 1: (3 kg−1, 3.5 bed−1) Site 2: (3 kg−1, 4 bed−1)	16	PSNR, NRMSE, SSIM, rSUV, CTD
[78]	Unet	3D Patch	None	18F-FDG	Brain	(5.18 kg−1, 5 bed−1)	7.5, 30	SNR, SSIM
[79]	CNN (Dilating convolutional kernels)	2D	None	18F-FDG	Brain	(166.5, 10)	10	MAE, PSNR, SSIM, rMAE
[80]	Unet	3D	None	18F-FDG	Brain	(205, 20)	20	PSNR, RMSE, SSIM, rSUV, QCS
[81]	FFNN	2D Patch	None	Sim 82Rb, 82Rb	Cardiac	(N/A, 7)	7, 3.5, 1.5	NMSE, ROI Contrast
[82]	Unet	3D Patch	None	18F-FDG	Whole body	(225.3, 10)	6.7, 9.1, 13.3, 17.5, 26.3, 66.7, 125, 250, 500	Lesion SUV, QCS, CTD
[83]	Modified Unet	2D	Sim T1	Sim 18F-FDG	Brain	(N/A, N/A)	N/A	MSE, Lesion CR
[84]	CNN	3D	T1	18F-FDG	Brain	(N/A, N/A)	10, 100	NRMSE, SUV bias
[85]	cycleGAN	2D Patch	None	18F-FDG	Brain	(218.3, 20)	125	PSNR, NRMSE, SSIM, SUV bias
[86]	GAN	2D	None	18F-FBB	Brain	(300, 20)	10	PSNR, NRMSE, SSIM, rSUV, QCS, CD
[87]	GAN	3D Patch	None	18F-FDG	Whole body	(5.55 kg−1, 20)	2	SSIM, PSNR
[88]	cycleGAN	3D Patch	None	18F-FDG	Whole body	BMI $\le$ 18.5: (370, 1.5 bed−1) $18.5\le$ BMI $\le$ 25: (370, 2 bed−1) $25\le$ BMI $\le$ 30: (370, 2.5 bed−1) 30 $\le$ BMI: (444, 2.5 bed−1)	8	MAE, NRMSE, rPSNR
[89]	cycleGAN	2D Patch	None	18F-FDG	Whole Body	(370, 5)	3.3, 10	PSNR, NRMSE SUV bias
[90]	GAN	2D Patch	None	18F-FDG	Whole body	(N/A, N/A)	10	PSNR, RMSE, SSIM Lesion SUV
[91]	GAN	2.5D	None	18F-FBB	Brain	(330, 20)	100	PSNR, RMSE, SSIM, FBM, EBM, CD
[92]	GAN	3D Patch	None	18F-FDG	Brain	(203, 12)	4	PSNR, nMSE, rSUV
[93]	GAN	3D Patch	T1, DT	18F-FDG	Brain	(203, 12)	4	PSNR, SSIM, rCR
[94]	GAN	3D Patch	None	18F-FDG	Whole body	(5.55 kg−1, 20)	5	NRMSE, PSNR, RFSIM, VIF
[95]	CAE, Unet, GAN	2D, 2.5D, 3D	None	18F-FDG	Thoracic	(370, 20)	10	PSNR, nMSE, Lesion SUV bias
[96]	Residual Unet	2D	T1, T2, FLAIR	18F-FBB	Brain	LD: (8, 30) FD: (334, 20)	42	PSNR, RMSE, SSIM rSUV, QCS, CD
[97]	Residual Unet	2D	T1, T2, FLAIR	18F-FBB	Brain	Site 1: (330, 20) Site 2: (283, 20)	Site 1: 100 Site 2: 20	PSNR, RMSE, SSIM rSUV, QCS, CD
[98]	Unet	3D Patch	None	18F-FDG, 18F-FMISO, 68Ga-Dotatate	Whole body	FDG: (340, 20) FMISO: (181, 50) DOTATATE: (130, 21.6)	10	PSNR, NRMSE, Lesion SUV bias
[99]	Residual Unet	2.5D	None	Sim 18F-FDG, 18F-FDG	Brain	(185, 70)	4	CR
[100]	Unet	2.5D	None	18F-FDG	Whole body	Site 1: (481, 3 bed−1) Site 2: (400, 3 bed−1) Site 3: (429, 3 bed−1)	4	QCS, CTD, rSUV
[101]	Residual Unet	3D	None	18F-FDG	Whole body	(391, 2.45 bed−1)	1.33, 2, 4	CTD, rSUV
[102]	Modified Unet	2.5D	T1, T2	18F-FDG	Brain	(230, 30)	180	PSNR, SSIM

DT diffusion tensor, PSNR peak signal-to-noise ratio, RMSE root mean square error, NRMSE normalised root mean square error, MSE mean square error, MAE mean absolute error, rSUV regional SUV, CR contrast recovery, SSIM structural similarity index, QCS qualitative clinical score, CD clinical diagnosis, CTD clinical tumour detectability, LVEF left ventricular ejection fraction, EDV end diastolic volume, ESV end systolic volume, LAVA liver acquisition volume acceleration, RFSIM Riesz-transform based feature similarity, VIF visual information fidelity, Sim simulated data