Room: Exhibit Hall | Forum 8
Purpose: Quantitative MR is clinically significant, since its signal is not influenced by imaging parameters. However, it is rarely applied in clinical practice, because it is typically derived from multiple images with slightly different weightings. Here we introduce a novel approach to derive a quantitative relaxation parametric map from a single qualitative image so that qualitative and quantitative MRI can be simultaneously obtained without changing standard imaging protocols.
Methods: We proved the concept in T_1 mapping. A series of T_1 weighted images were acquired using a 3D ultra-short echo time (UTE) sequence with variable flip angles. T_1 map was derived using non-linear fitting.Deep learning techniques were used to incorporate a priori information of relaxation properties into the derivation of quantitative MRI. A non-local deep neural network was established to provide a mapping from a single T_1 weighted image to the corresponding T_1 map. The network had a hierarchical architecture with shortcut connections established. Additionally, global information was incorporated via the use of the attention mechanism, which could capture long-range dependencies across image regions. An attention layer was integrated to every convolutional block, where signal at a position was obtained by attending to all positions in the previous feature map. Several flip angles were applied to establish different prediction models, each trained and tested using images acquired at a specific flip angle. We used 1224 image for training and 192 images for testing.
Results: The predicted T_1 maps had high fidelity to the ground truth maps that were measured using variable flip angles. The averaged correlation coefficients were 0.9542, 0.9626, 0.9728 and 0.9742 when flip angle was 5Â°, 10Â°, 20Â° and 30Â°.
Conclusion: A deep learning based approach was developed to derive quantitative MRI from a single qualitative MRI. Therefore, qualitative and quantitative MRI can be simultaneously acquired.