The feasibility of water-fat separation using an end-to-end ConvNet approach was demonstrated for complex, magnitude and single echo acquisitions. The ConvNet approach showed images visually comparable to the GraphCut method with slightly higher signal to noise in typical cardiac image planes. Quantitative PDFF, R2* and off-resonance values had excellent correlation with a conventional analytical model based method. ConvNet based water-fat separation is a promising method capable of learning the water-fat separation problem with corrections for bipolar gradients, a multi-peak model, R2* and off-resonance.
This retrospective study used complex raw data from a database of prospective research cardiovascular studies. The database contained relevant studies from 66 subjects: 17 subjects without a cardiovascular history or complaints (normal controls), 13 acute myocardial infarction (MI) subjects at 3 days after infarction and 34 chronic MI subjects (>2 years after infarction). Additionally, there were 26 repeat sessions with the acute MI patients in the subacute phase of infarction. In total, 90 imaging session were included. The database consisted of 1204 acquisitions.
All MR examinations were performed using one of two clinical 1.5 T imagers. The pulse sequence was a dark blood double inversion recovery multiple spoiled gradient-echo sequence (1 slice per breathhold, repetition time = 20 ms; 12 echo times, 2.4 - 15.5 ms (1.2 ms spacing), flip angle = 20 degrees, bandwidth = 1860 Hz/pixel, in-plane spatial resolution = 2.3 x 1.7 mm, slice thickness = 8 mm, flow compensation in read and slice). Every subject had acquisitions in 2-, 3- and 4- chamber long axis as well as short axis planes spanning the left ventricle. Multi-channel complex raw data was saved to the scanner’s hard drive and archived. Conventional Method Conventional water fat separation was performed via a multi-point fat-water separation with R2* using a graph cut field map estimation algorithm (1) with the ISMRM water-fat Toolbox (2). Conventional processing of the bipolar acquisition was done separately using the even and odd echo images and results then averaged, producing water, fat, R2* and off-resonance images for each acquisition.
120 images were identified with artifacts and excluded from training. Six imaging sessions, one normal, three chronic and two acute MI patients consisting of 80 acquisitions were also held out for method evaluation and not used in the training process. The images were used for ConvNet prediction and subsequent quantitative analysis of water- and fat-only images, R2* and off-resonance. 1000 acquisitions were then available for training (900 fitting and 100 validation).
A U-Net (3) convolutional neural network was used for deep learning water-fat separation (Figure 1). Using all available data, the input to the training algorithm was 24 real and imaginary images from 12 echo times used as 24 channels of the ConvNet. Also investigated was the use of magnitude only inputs. Additionally, single echo acquisitions were simulated using only the first opposed-phase echo (TE=2.4 ms). The implementation was realized using Keras 2.0 (30) and TensorFlow 1.2 (31) (both freely available software). An EVGA GEFORCE GTX 1080 Ti graphics card was used for GPU accelerated calculations.