A deep learning approach for synthetic MRI based on two routine sequences and training with synthetic data

• Development of a learning-based approach to compute T1, T2, and PD maps from clinical routine sequences. Computation three realistic parametric only two input images (a three-from-two approach). Realistic are obtained actual data by training the network with synthetic dataset. weighted pertaining modalities unseen maps. Quantitative MRI in viable times. Background Objective: Synthetic magnetic resonance imaging (MRI) is low cost procedure that serves as bridge between qualitative quantitative MRI. However, proposed methods require very specific sequences or private protocols which have scarcely found integration scanners. We propose pair T1- T2-weighted customarily acquired routine. Methods: Our based on convolutional neural (CNN) trained data; specifically, dataset 120 volumes was constructed anatomical brain model BrainWeb tool served set. The CNN learns an end-to-end mapping function transform their underlying Then, conventional analytically synthesized can be fine tuned small database for better performance. Results: This able accurately achieving normalized squared error values predominantly below 1%. It also yields MR acquisitions range literature correlation above 0.95 compared T1 T2 relaxometry Further, visually realistic; mean square always 9% structural similarity index usually 0.90. Network tuning improves performance, while exclusively shows performance degradation. Conclusions: These results show our provide out (a) (b) needs inputs, turn common full-brain acquisition takes less than 8 min scan time. Although crucial reach acceptable levels. Hence, we utility both times synthesis additional those actually acquired.