. 2018 Nov 20;46(1):e1–e36. doi: 10.1002/mp.13264

Table 6.

Image processing and reconstruction with DL

Task	Imaging modality	Performance measure	Network output	Network architecture basis
Filtering	CT,²³⁴Chest x ray,²³⁵x ray fluoro²³⁶	MSE²³⁴, CAD performance,²³⁴PSNR,²³⁵^,²³⁶ SSIM,²³⁵^,²³⁶Runtime²³⁶	Likelihood of nodule,²³⁴Bone image,²³⁵CLAHE filtering²³⁶	Custom CNN,²³⁴^,²³⁵Residual CNN,²³⁶Residual AE²³⁶
Noise reduction	CT,²³⁷^,²³⁸^,²³⁹^,²⁴⁰PET²⁴¹	PSNR,²³⁷^,²³⁸^,²³⁹^,²⁴⁰^,²⁴¹ RMSE,²³⁷^,²³⁸ SSIM,²³⁷^,²³⁸^,²⁴⁰ NRMSE,²³⁹ NMSE²⁴¹	Noise‐reduced image²³⁷^,²³⁸^,²³⁹^,²⁴⁰^,²⁴¹	Custom CNN,²³⁷^,²³⁸^,²³⁹Residual AE,²³⁷^,²³⁸Concatenated CNNs,²⁴¹U‐net²⁴⁰
Artifact reduction	CT,²⁴²^,²⁴³MRI²⁴⁴	SNR,²⁴²^,²⁴³ NMSE,²⁴⁴ Qualitative,²⁴³ Runtime²⁴⁴	Sparse‐view recon,²⁴²^,²⁴⁴ Metal artifact reduced image²⁴³	U‐net,²⁴²^,²⁴⁴Custom CNN²⁴³
Recons	MRI²⁴⁵^,²⁴⁶^,²⁴⁷^,²⁴⁸	RMSE,²⁴⁵^,²⁴⁸ runtime,²⁴⁵ MSE,²⁴⁶^,²⁴⁷ NRMSE,²⁴⁶ SSIM,²⁴⁶ SNR²⁴⁸	Image of scalar measures,²⁴⁵ MR reconstruction²⁴⁶^,²⁴⁷^,²⁴⁸	Custom CNN,²⁴⁵^,²⁴⁸Custom NN,²⁴⁶Cascade of CNNs²⁴⁷
Registration	MRI²⁴⁹^,²⁵⁰^,²⁵¹^,²⁵²x‐ray to 3D²⁵³^,²⁵⁴	DICE,²⁴⁹^,²⁵⁰ Runtime,²⁵⁰ Target Overlap,²⁵¹ SNR,²⁵² TRE,²⁵⁴ Image and vessel sharpness,²⁵² mTREproj²⁵³	Deformable registration,²⁴⁹^,²⁵⁰^,²⁵¹^,²⁵² Rigid body 3D transformation²⁵³^,²⁵⁴	Custom CNN,²⁴⁹^,²⁵¹^,²⁵²^,²⁵³^,²⁵⁴SAE²⁵⁰
Synthesis of one modality from another	CT from MRI,²⁵⁵^,²⁵⁶^,²⁵⁷^,²⁵⁸^,²⁵⁹MRI from PET,²⁶⁰PET from CT²⁶¹	MAE,²⁵⁵^,²⁵⁶ PSNR,²⁵⁵^,²⁵⁹ ME,²⁵⁶ MSE,²⁵⁶ Pearson correl,²⁵⁶ PET image Quality,²⁵⁷^,²⁵⁸ SSIM,²⁶⁰ SUVR of MR‐less methods,²⁶⁰ Tumor detection by radiologist²⁶¹	Synthetic CT,²⁵⁵^,²⁵⁶^,²⁵⁷^,²⁵⁸Synthetic MRI,²⁶⁰Synthetic PET²⁶¹	Custom 3D FCN,²⁵⁵GAN,²⁵⁹^,²⁶⁰^,²⁶¹U‐net,²⁵⁶^,²⁵⁷AE²⁵⁸
Image quality assessment	US,²⁶² CT,²⁶³^,²⁶⁴ MRI²⁶⁵	AUC,²⁶²^,²⁶⁴ IOU,²⁶² Correlation between TRE estimation and ground truth,²⁶³ Concordance with readers²⁶⁵	ROI localization and classification,²⁶² TRE estimation,²⁶³ Estimate of image diagnostic value²⁶⁴^,²⁶⁵	Custom CNN,²⁶²^,²⁶⁵Custom NN,²⁶³VGG19²⁶⁴

MSE, mean‐squared error; RMSE, Root MSE; NSME, normalized MSE; NRMSE, normalized RMSE; SNR, signal‐to‐noise ratio; PSNR, peak SNR; SSIM, structural similarity; DICE, segmentation overlap index; TRE, target registration error; mTREproj, mean TRE in projection direction; MAE, mean absolute error; ME, mean error; SUVR, standardized uptake value ratio; AUC, area under the receiver operating characteristic curve; IOU, intersection over union; CLAHE, contrast‐limited adaptive histogram equalization.