Exercise‐induced calf muscle hyperemia: Rapid mapping of magnetic resonance imaging using deep learning approach

Figures

• 

  Figure 1. Schematic diagram of perfusion mapping in dynamic contrast enhanced magnetic resonance imaging (DCE‐MRI). In the acquired dynamic images, an arterial input function (AIF) is manually sampled from a large arterial region, and all the images are converted to maps of tissue contrast enhancement (TC). For each voxel of tissue, parameter values of impulse retention function (IRF) are optimized to fit the convolution of AIF and IRF to the voxel's TC. Tissue perfusion is a parameter of IRF. Completion of the model fitting for all voxels in a slice would result in a perfusion map. Dimensions of each data are shown in parentheses. For example, AIF is a one‐dimension vector as a function of time, and perfusion map has the same dimension as image slice

• 

  Figure 2. Estimation error of the different methods, over the range of reference perfusion. Along the X axis, the reference values were discretized by intervals of 10 ml min−1 100 g−1. For all reference values in each interval, the errors of perfusion estimates by a method were averaged and set as the Y‐axis value of the plotted point. The low‐error spike near to perfusion of 300 ml min−1 100 g−1 in the F1 (regular model fitting) curve was due to misfit for a group of voxels by model fitting. For these voxels, optimization of both the multigrid fit and regular model fit left the parameter of perfusion unchanged at its initial values of 300 and 350 respectively. This overestimation by F1 at 300 reduced its averaged error over the perfusion interval of 290–300 ml min−1 100 g−1. F1 and N1‐4 denotes the different methods as specified in the caption of Table 2

• 

  Figure 3. Histogram of perfusion values in the training and testing groups. Median and mean for the diverse training group (20 datasets) were 80.8 and 106.6, for the homogeneous training group (20 datasets) were 62.8 and 91.9, and for the testing group (eight datasets) were 72.4 and 103.8 ml min−1 100 g−1. The spikes at perfusion value of 300 ml min−1 100 g−1 were because for multiple voxels presumably with high perfusion, the model fitting ended at a local optimum with the initially chosen perfusion value unchanged at 300

• 

  Figure 4. Perfusion maps generated by the various methods for a same subject data. (a) Perfusion map by the multigrid model fitting (F0); (b) regular model fitting (F1); (c) NN trained by 20 diverse data (N1). All the maps were cropped to remove most background regions