Background Water–fat separation is a postprocessing technique most commonly applied to multiple‐gradient‐echo magnetic resonance (MR) images to identify fat, provide images with fat suppression, and to measure fat tissue concentration. Recently, Numerous advancements have been reported. In contrast to early methods, the process of water–fat separation has become complicated due to multiparametric analytic models, optimization methods, and the absence of a unified framework for diverse source data. Purpose To determine the feasibility and performance of MRI water–fat separation and parametric mapping via deep learning (DL) with a range of inputs. Study Type Retrospective data usage. Population/Subjects Ninety cardiac MR examinations from normal control, acute, subacute, and chronic myocardial infarction subjects were obtained, providing 1200 multiple gradient‐echo acquisitions. Field Strength/Sequence 1.5 T/2D multiple gradient‐echo pulse sequence Assessment Ground‐truth training and validation water–fat separation were obtained using a graph cut method with R2*, off‐resonance correction, and a multipeak fat spectrum. U‐Net DL training with single and multiecho, complex, and magnitude inputs were compared using quantitative and three‐observer subjective analysis. Statistical Tests DL methods' image structural similarity, and quantitative proton density fat fraction (PDFF), R2*, and off‐resonance quantitative values were statistically compared with the GraphCut reference standard using Student's t‐test and Pearson's correlation. Results Myocardial fat deposition in chronic myocardial infarction and intramyocardial hemorrhage in acute myocardial infarction were well visualized in the DL results. Predicted values for R2*, off‐resonance, water, and fat signal intensities were well correlated with a conventional model‐based water fat separation (R2 ≥ 0.97, P < 0.001) with appropriate inputs. DL parameter maps had a 14% higher signal‐to‐noise ratio (P < 0.001) when compared with a conventional method. Data Conclusion DL water–fat separation is feasible with a wide range of inputs, while R2* and off‐resonance mapping requires multiple echoes and complex images. With appropriate inputs, DL provides quantitative and subjective results comparable to conventional model‐based methods. Level of Evidence: 1 Technical Efficacy Stage: 1 J. Magn. Reson. Imaging 2019;50:655–665.

We investigate the effectiveness of a simple solution to the common problem of deep learning in medical image analysis with limited quantities of labeled training data. The underlying idea is to assign artificial labels to abundantly available unlabeled medical images and, through a process known as surrogate supervision, pre-train a deep neural network model for the target medical image analysis task lacking sufficient labeled training data. In particular, we employ 3 surrogate supervision schemes, namely rotation, reconstruction, and colorization, in 4 different medical imaging applications representing classification and segmentation for both 2D and 3D medical images. 3 key findings emerge from our research: 1) pre-training with surrogate supervision is effective for small training sets; 2) deep models trained from initial weights pre-trained through surrogate supervision outperform the same models when trained from scratch, suggesting that pretraining with surrogate supervision should be considered prior to training any deep 3D models; 3) pre-training models in the medical domain with surrogate supervision is more effective than transfer learning from an unrelated domain (e.g., natural images), indicating the practical value of abundant unlabeled medical image data.

No abstract

No abstract