ausblenden:
Schlagwörter:
-
Zusammenfassung:
INTRODUCTION: To make 7T CEST MRI more available for radiologists, we developed a deepCEST pipeline for 7T MRI that predicts CEST contrasts from just one scan with robustness against B1 inhomogeneities. The pipeline includes an uncertainty quantification and a confidence map to directly evaluate the quality of the predictions. The proposed approach results in a reduction of 50% of the measurement time and delivers the predicted CEST contrast with in 1 sec. METHODS: The input data for a neural network (NN) consisted of 7T in vivo raw Z-spectra of a single B1 level (1 μT), and B1 and B1-CP maps. The 7T raw data was acquired using the 3D snapshot GRE MIMOSA CEST3 at a Siemens MAGNETOM 7T scanner. These inputs were mapped voxel-wise on target data consisting of Lorentzian amplitudes conventionally generated by 5-pool-Lorentzian fitting performed on normalized, denoised, B0- and B1-corrected Z-spectra1. The network consisted of two fully connected hidden layers with RELU activation and was trained with Gaussian negative log likelihood loss. The output layer consisted of 10 nodes with linear activation to obtain amplitudes and log variance of the 5-pool Lorentzian fit. Exponential activation was applied to the log variance during inference to obtain the uncertainties. RESULTS: Figure 1 shows the Lorentzian fit and the prediction of the amide and rNOE contrast in a tumor patient with a B0 shift, as well as the segmented uncertainty map over all contrasts with a threshold of 5%. The first row (Fig. 1a) shows the predictions and uncertainty for the measurement made with a bad shim. Such a strong B0 shift was not part of the training distribution. Therefore, the predictions and fits do not match. Consequently, the NN outputs a high uncertainty for these voxels. In the second row (Fig. 1b), the Z-spectra of the patient were centered, resulting in NN predictions that agree well with the fit, show the tumor highlighted, and only a low uncertainty is yielded. In figure 1c the hisogram of the B0 map from case a and b is shown, in which the strong B0 shift is shown. The B1 homogeneity of the deepCEST prediction, inferred from only one B1 level (1 μT) is as good as the fit, generated by B1-interpolation of 0.72 μT and 1 μT CEST data4. Predicted maps were also observed to be more homogeneous and smoother owed to the denoising ability of the deepCEST networks. DISCUSSION and CONCLUSION: The deepCEST approach has already shown very promising results for 3T, the clear advantage of 7T data is the better SNR and higher spectral resolution. The 7T deepCEST approach uses only one B1 level, this saves about 50% of scan time (now 6:22 min), but still predicts accurately with low uncertainty and provides both B0- and B1-corrected homogeneous CEST contrast together with an uncertainty map, increasing the diagnostic confidence.
